Cerebellum

[rcoreilly]

It had to happen eventually.. If we want to actually make a model with any kind of reasonable sensory-motor control abilities, it has to have a cerebellum. Recent attempts to understand the neural pathways for representations of Space #332 have bumped into this issue, because it is entirely clear that the core head direction circuit (HD) involves a ring attractor system between the DTN (dorsal tegmental nucleus of Gudden) and the AD (anterior-dorsal nucleus of the thalamus). Such a system is inherently unstable and requires precise weight adjustments, and it is most likely that this comes from the cerebellum. Certainly there is abundant evidence that the inputs to this system from the vestibular nuclei are based on multimodal integrated sensory signals that combine vestibular, motor efference and proprioception, and visual optic flow, which are known to be tuned by the cerebellum (Cullen19, Cullen23).

Somehow, the literature on all this appears to be remarkably fragmented, with no dominant computational model of cerebellar function being discussed in the empirical papers on the topic (e.g., the Cullen papers). Cullen instead cites LaurensAngelaki18 who have a Kahlman filter model (LaurensAngelaki17), which makes all the right points about internal models and error signals in interpreting the outputs of the vestibular system, but does not provide any kind of actual circuit mapping of the operation of the proposed Kahlman filter onto the actual cerebellum. Major review papers like KheradmandZee11 on cerebellum and motor control provide compelling evidence of all the good things it does, without any citations to a computational model!

[rcoreilly]

Recent reviews

NguyenPerson25

good place to start it seems..

Although often considered a particularly ancient structure, the cerebellum proper first appeared in jawed vertebrates, well after pallial (cortex), basal ganglia, brainstem and spinal circuits evolved 2.

According to current understanding, its emergence coincided evolu- tionarily with a massive expansion of vertebrate body size and motor repertoires. The evolution of larger bodies led to two novel motor con- trol problems that the cerebellum has been proposed to solve. First, it made sensory feedback delays longer as larger bodies increase axonal conductance times between the periphery and the brain. Without the emergence of novel computations to handle such delays, movement corrections reliant on sensory feedback would therefore be based on out-of-date information. Second, larger bodies resulted in more complex self-generated sensory feedback and interaction torques — forces that arise from the action of one body part on another — that the brain must mitigate for effective control. Interaction torques scale multiplicatively with the weight and length of the effector: thus, from a control theory standpoint, the sharp increase in self-generated forces as bodies grew larger necessitated a neural system with the computa- tional capacity to compensate for (through predictive computations) and interpret (as self-generated rather than non-self-generated) these forces3

Key point: forward model provides a defined source of error signals, while this is unclear in FF model.

How might the cerebellum solve these motor control problems? Some theories propose that the cerebellum computes internal models, in which sensory predictions are generated from motor commands to shorten feedback delays through internal simulation. Internal models serve many purposes, and could, in principle, solve both the sensory delay problem and self-generated versus non-self-generated feedback problems described above5,7–10. Another class of solutions, however, sidesteps internal simulation and instead posits that the cerebel- lum serves a more basic function, that of feedforward anticipatory control11–13.

This Review explores whether the cerebellum’s well-understood role in associative learning can account for its contribution to motor control. We discuss the two competing, but not mutually exclusive, ideas described above — that the cerebellum computes feedforward control policies, generating predictive motor commands to influ- ence movements, or, alternatively, that it computes forward internal models, generating sensory predictions that are used in diverse ways by downstream structures (Fig. 2). This latter function is well studied in a canonical cerebel- lar associative learning paradigm, delay eyelid conditioning (DEC).

Our discussion focuses on multiple stages along the cerebellar circuit, piecing together a layer-by-layer algorithm that produces canonical predictive DEC and comparing and contrasting this algorithmic solution with the signals generated dur- ing movements. Finally, we end with a discussion of how this control solution interfaces conceptually with feedback control using internal models and evaluate whether, on balance, the current state of the field favours either model. The scope of this Review focuses mainly on relating DEC to other discrete movements, and thus we do not com- prehensively include all cerebellar adaptation paradigms, including vestibular learning and control.

Ooops, maybe not so relevant.

by reweighting of the subset of parallel fibre synaptic inputs onto a PC that are active just before the air puff, with this reweighting being driven by an instructive signal from the climbing fibres that conveys air puff information and drives a complex spike in PCs17,22. This results in a reduction in the excitatory parallel fibre-to-PC drive that, coupled with stable feedforward inhibition from molecular layer interneurons (MLIs), pauses PC firing around the time of the instructive complex spike23–26. At the core of this computation is the translation of an esti- mate of the elapsed time between the CS and the US into a movement, in this case an eyelid closure that occurs before a reflexive blink would otherwise need to happen.

DEC reveals a form of feedforward motor control that links an early cue with a later response through associative learning. This form of control could explain cerebellar contributions to many diverse movements11,12,17, for each of which within-movement contextual informa- tion could be used in ways analogous to the CS in DEC.

Whereas DEC provides an example of learned anticipatory feed- forward control by the cerebellum, the alternative hypothesis that the cerebellum computes internal models used for rapid error correction is the focus of another dynamic subfield of cerebellar research5,8,9,30. For- ward internal models, which predict sensory consequences of actions from motor efference copies, could be used to generate anticipatory corrections5,9,31–33, but to do so they require a secondary calculation, sensory prediction, that is fully bypassed in the feedforward con- trol framework. A second form of internal model, inverse models, in which motor commands are computed to accomplish a motor goal, also require extensive computation as they must model joint torques that vary wildly under different conditions8,34. By contrast, feedfor- ward control policies simply associate the sensorimotor conditions (efference copy, sensory feedback and goal) with the adjustments that the brain had to make in the past under such conditions, and generate those adjustments pre-emptively as a learned anticipatory command11,35,36.

Despite these findings, individual PC simple spike patterns remain
puzzling. Their response patterns are highly diverse, even among PCs
with shared complex spike tuning (that is, PCs that respond to errors
in a common direction), and the activity of single neurons accounts
for relatively little behavioural variance80,85. This raised the question
of whether additional types of coding by PCs contribute to cerebellar
output80. A breakthrough came from studies showing that PC tuning
diversity becomes more interpretable when these cells are considered
at the population level, with behavioural tuning strengthening as analyses incorporate larger PC ensembles80,85. Investigators initially inferred that PCs with shared complex spike tuning are more likely to converge anatomically within the cerebellar nuclei, and therefore pooled these PCs to determine their encoding as a population. This pooling significantly improved behavioural encoding accuracy80,85. Notably, similar results have been observed even in studies in which complex spike tuning was not directly assessed — that is, when con- vergence was instead inferred through the topographic projections of PCs onto the cerebellar nuclei61,84. In summary, groups of PCs more closely align with kinematic variables, consistent with population coding hypotheses.

This pattern is indeed observed in PC populations that control saccades, reaches, and tongue and whisker movements, with net suppression of the PC population occurring at discrete moments during movements61,84,86–88.

The rapid kinetics of PC–nuclear interactions allow nuclear neurons to respond within mil- liseconds to population disinhibition, with this disinhibition driven by the precise temporal alignment of PC spikes and pauses91,92.

Some reports suggest that facilitating PCs belong to upbound modules, which are zebrin-positive zones that receive instruc- tive input from specific regions of the inferior olive, while suppressing PCs correspond to downbound, zebrin-negative zones94. A recent study that modulated two of these channels (one upbound and one down- bound) independently showed that they clearly actuate distinct types of eye movements94. This suggested that at least these neighbouring modules did not obviously converge and collaborate94, which calls into question the interpretations around population codes discussed above, which are based on merged facilitating and suppressing PCs80,84.

For example, a recent preprint has reported that a sizable population of PCs show simple spike facilitation that is dependent on pausing PCs disinhibiting downstream cerebellar outputs that, in turn, engage feedback networks to excite PCs95. These facilitating PCs may contribute to agonist/antagonist muscle coordination and timing95.

Moreover, our emerging appreciation of the molecular diversity of PCs, which cluster into multiple groups based on transcriptome expression, could lead to identification of distinct computational modules that intermingle within the cerebellar cortex94,100. Already, studies support the idea that there are links between molecular and physiological diversity, some of which map onto the traditionally appreciated zebrin-positive and negative zones94,101. Thus, future work incorporating connectivity, molecular and physiological diversity and computational hypotheses of cerebellar contributions to motor control promises to yield rapid progress in resolving these debates in upcoming years.

Climbing fibres also signal sensorimotor error in various motor adaptation paradigms, such as vestibulo-ocular reflex gain and smooth pursuit adaptation, in which climbing fibres signal visual motion detected on the retina when the visual target is supposed to be stabilized by eye movement itself105–107, consistent with a role for these signals in tuning internal models. Similarly, saccades that land too far to one side of a target will elicit complex spikes81,108,109 that are tuned to the direction of the error but in sensory coordinates, such as the location of a visual target relative to the fovea109, as predicted by the forward model hypothesis109. The PCs that receive this signal are those whose simple spike firing rates must be suppressed to pull the eyes to the required position. Thus, it is proposed that the climbing fibres provide an instructive signal that will induce plasticity that serves to reduce the firing of this population of PCs, leading to an adjustment in the saccade end-point in subsequent saccades81.

Despite the dominance of the error signal hypothesis, the canoni- cal error signals that it predicts do not align with many observations of complex spikes in the literature. Interested readers are referred to excellent recent reviews focused on complex spike tuning diversity110,111.

The most salient complex spike signals during monkey reaches occur before the movement112 and complex spike tuning has been shown to predict the movement end-point, rather than responding to end-point errors112. Similarly, in monkeys using a manipulandum to track a randomly mov- ing dot, complex spikes predicted errors, rather than responding to them113,114. This experimental structure deprives the subject from relia- ble sensory predictors with which to generate an associative correction, thus the finding that these complex spikes predict errors is equivalent to saying that they lead errors, probably by perturbing simple spiking to causally influence movements. Indeed, the view that complex spikes can drive movement adjustments, either in advance of or in response to sensory events, has been proposed for other behaviours, including saccades108. This perturbation of simple spikes by complex spikes has also been suggested to serve an important role in learning, through an instruct-then-reinforce role115.

One proposed explanation for these observations is provided by the idea that complex spike tuning shifts to encode subjective salience, which prepares action states117,118.

Of course, the tuning of complex spikes is of interest mainly because these signals drive plasticity at the parallel fibre-to-PC syn- apse, resulting in motor learning38,106,107,130–134. Coincident signalling of climbing fibres with parallel fibres leads to long-term depression of these synapses whereas parallel fibre firing alone leads to long-term potentiation115,135. This elegant, bidirectional learning rule forms the basis of powerful computations underlying the cerebellum’s proposed role as an adaptive filter36,136. For example, if PC simple spike rates are monitored over multiple trials in a motor learning task, it is seen that rates are lower for trials that follow a trial in which a complex spike was generated and higher for trials that follow a trial in which there was no complex spike, demonstrating a rapid form of learning81,84,106,107,133. Moreover, pharmacological or optogenetic manipulation of the inferior olive to drive complex spike rates up or down can incept associations or block the extinction of associations, respectively, observations that are also remarkably consistent with these plasticity rules102–104,132,134.

An intriguing insight that could potentially explain this heterogeneity comes from the discovery that regional diversity exists in the temporal associativity rule inducing parallel fibre long-term depression: although coincident activation of climbing fibres and parallel fibres drove plasticity in vermal regions, it failed to induce synaptic depression in a distinct cerebellar region, the flocculus137. There, synaptic depression was instead reliably induced when parallel fibre activity preceded climbing fibre activity by 120 ms, which is also the predicted latency of visual feedback to the structure. A recent preprint has reported that the length of the instructive win- dow is under learned control itself138, responsive to naturalistic error feedback delays for different cerebellar control modules. Thus, the timing of feedback signals may be linked to the eligibility trace for synaptic learning rules.

Because MLI-mediated inhibition of PCs can block associative learning142,144 and enhanced excitation of PCs can boost associative learning, it will be important to investigate whether the diversity of MLI motifs can explain how learning is gated, protected and expressed.

n22,40 (Fig. 5). This population coding motif solves the problem of learning elapsed time between stimuli by facilitating the temporal overlap of late-activating GCs and the inferior olive-driven complex spike. This means that plastic- ity is induced only at synapses made by GCs active at the end of the CS interval, sparing the synapses made by all other active GCs.

In 2017, a bevy of papers reported the outcomes of experiments and theories probing the Marr and Albus prediction of GC layer population activity sparseness150–154. Disagreement initially arose around the question of whether the pre- dicted GC sparseness matched the in vivo data.

However, the current consensus is that statistical reformatting of mossy fibre information by the GC layer, resulting in decorrelation and temporal expansion (resulting in a more granular representation of time-evolving activity), can enhance learning, even without extensive sparsening155–157. Additionally, when the behaviour under investigation is higher dimensional (involving many body parts simultaneously) and faster this appears to increase GC population information dimensionality155, suggesting that evidence of low-dimensional, dense activation could reflect task demands, rather than the coding principle at work.

Calcium imaging and serial single-cell recordings from GCs lend sup- port to the notion that a time-evolving population code exists, par- ticularly during limb movements in mice and following electric-organ discharges in electric fish148,156. Recordings from GC populations in mice during tasks with varying delay intervals between operant movement completion and reward delivery showed that the GC population activ- ity both tiled time, with unique subpopulations active at evolving time intervals, and flexibly matched the temporal intervals of the task156.

On the other hand, during tasks involving sensory stimuli (such as the tones used in DEC), GC activity timing was either very dense and non-stationary over learning151 or inconsistent across trials157, neither of which is easily reconciled with the temporal basis set hypothesis.

GC excitability gradients are also observed across the dorsal–ventral axis of the layer: thus, if identical inputs were shared across that axis, timing diversity could emerge simply through excitability differences43,170.

Moreover, the idea that sensory prediction and sensory prediction errors are encoded by the cerebellum — either as PC codes80,122 or as cancelled sensory reafference in the cerebellar nuclei51,61,89 — is supported mechanistically.

[rcoreilly]

Initial hypotheses

Here's some initial thoughts:

1. Everyone is overestimating the contribution of the cerebellar cortex and underestimating the role of the DCN, which may in fact be doing the most important work.
2. All the basic cortical GC -> PC stuff seems reasonable, but it is essential to understand what drives the IO if you're going to understand anything about learning. IO drives simple spikes and complex spikes. Simple and complex spikes encode a heck of a lot of information: errors, kinematics, etc.
3. IO gap junction coupling is controlled by DCN.. everything is controlled by DCN.

Round 2 hypotheses

• per Herzfeld et al 21: DCN and CerCtx are two parallel but coordinated learning systems.
• DCN -> IO has excitatory and inhibitory pathways that can compute the prediction error.
  • In vestibular / old cerebellum, this is available directly, but in more modern cerebellum, the excitatory part goes through the MDJ, which also gets neocortical and other inputs.
  • In all cases, the prediction is DCN inhib -> IO, which is generally thought to be synonymous with the current control output of Cereb.
  • Excitatory are the sensory or motor outcomes. Makes sense that this includes neocortex and other inputs in MDJ.
• Each DCN is specialized for sensory or motor predictions? needs to coordinate prediction and outcome signals.
• IO Gap junction mech allows sensor or motor errors to activate learning signal for everyone!! perfect!
• Intrinsic oscillation in IO provides temporal window for prediction - outcome offsets -- prediction comes first as a DCN motor control / modulatory signal, then the sensory / motor efferent outcome comes ~50 ms later or so (characteristic timing built-in). Only if it comes outside of the prediction window does the IO fire.
• It is important that the excitatory inputs to DCN are also the same that go to Cereb. Cortex so they can learn the same things.
• presumably DCN inhib plasticity is on same inputs..
• IO-driven CS establishes one-trial quick learning that then provides ongoing learning signal to DCN to "consolidate", per Herzfeld.

TODO: make a figure, then make a model of a simple vestibular case. so sensible that this vestibular thing is the most ancient part of the cerebellum..

Round 3:

• Started to question the independent CN learning framework. Are there even strong enough CNinhib -> CNexcite projections? Some conclusions were pretty weak: UusisaariKnopfel08 KebschullCasoniConsalezEtAl24 -- however, going back to the original deZeeuw PhD thesis showed that there are as many non-PC inhibitory synapses as PC ones.
• So much data consistent with independent function in CN after cerebellum lesions.
• OTOH, the ZobeiriCullen24 and TanakaIshikawaKakei19 data clearly show that PC inhibits DCN.
• And lots of data really does support idea that PC needed for learning in CN. This is tough to reconcile with an independent learning process in CN. The GarciaSteeleMauk99 result shows no learning after PC lesions, for DEC. But that could be weird. But KassardjianTanChungEtAl05 showed no CN learning in VOR, which is much more basic and doesn't involve delays etc. Need to look into this more carefully.
• So whereas it seems clear that CNi -o CNe inhibition can support the learned result, maybe PC input is actually required for learning? Herzfeld et al also have this key feature, needed to explain their data.

[rcoreilly]

DCN, IO circuitry

DeZeeuwSimpsonHoogenraadEtAl98

The DCN send both inhibitory and excitatory connections onto IO, contained within glomeruli: how could this NOT be the locus of an error-signal computation as the difference between the two?

DeGruijlBosmanDeZeeuwEtAl13

The definitive reference on IO

In general, a glomerulus contains a core of five to ten dendritic and/or axonal spiny appendages derived from different neurons (De Zeeuw 1990; De Zeeuw et al. 1990b, c). This core is surrounded by four or five terminals and several glial sheaths. Serial section analysis has demonstrated that virtually all spines are located in glomeruli.

Instead, with multiple unit recordings following application of harmaline and picrotoxin, Llina ́s and colleagues have demonstrated that in the intact IO, synchronous firing can be induced in coupled cellular aggre- gates of hundreds of neurons (Llina ́s and Volkind 1973; Lang et al. 1996). In fact, bilateral multiple unit recordings from the cerebellar cortex demonstrated that an ensemble of coupled IO neurons in the rat can even extend beyond the midline (De Zeeuw et al. 1996). In the same study, it was estimated that one IO neuron may have 500–1,000 gap junctions, and that two individual IO neurons may be coupled by 10–20 gap junctions (De Zeeuw et al. 1996). Estimates of the density of neuronal gap junctions with the use of anti-Cx36 are in line with these counts (Fig. 43.4a).

The fact that the density of dendritic lamellar bodies in the IO is higher than in any other area of the brain (De Zeeuw et al. 1995) points to the importance of electrotonic coupling between IO neurons.

The IO receives sensory and motor signals from nuclei such as the trigeminal nuclei, dorsal column nuclei, pretectal complexes, and red nuclei as well as direct feedback signals from DCN and feedback signals that are relayed via the mesodiencephalic junction (Swenson 1983; De Zeeuw and Ruigrok 1994; Onodera and Hicks 1995).

In the PO and MAO, all terminals derived from the mesodiencephalic junction are excitatory and display the corresponding morphological features consisting of rounded vesicles and asymmetric synapses. In contrast, all the cerebellar terminals in the PO and MAO are GABAergic and have pleiomorphic vesicles and symmetric synapses (De Zeeuw et al. 1988). Approximately half of both types of terminals contact dendritic elements inside glomeruli (De Zeeuw et al. 1989, 1990b, c, d; Strata 1989; De Zeeuw 1990).

The large majority of the remaining terminals contact the proximal and intermediate dendrites, while relatively few terminate on the somata and axon hillock; presynaptic axo-axonal contacts have not been observed in the IO. The innervation of the IO by the non-GABAergic mesodien- cephalic terminals and GABAergic cerebellar terminals is apparently random, because neither type of terminal has a preference for either the extra- or intraglomerular neuropil and there is no obvious pattern in the distribution of the two types of terminals within the individual glomeruli (Strata 1989). Every spine on the dendrites and axon hillock of all IO neurons in the PO and rostral MAO receives a synaptic input from both an excitatory mesodiencephalic terminal and an inhibitory cerebellar terminal. Since in most regions of the central nervous system, the vast majority of dendritic spines are contacted solely by asymmetric synapses, the ubiq- uitous, combined excitatory and inhibitory input to the IO spines can also be considered as one of the characteristic features of its neuropil. The other subnuclei such as the DAO, b nucleus, DCK, VLO, and DMCC follow the same configuration, but with different origins of the inputs involved (De Zeeuw et al. 1993, 1994, 1996).

IO combination of "paired" excitatory and inhibitory inputs is "unique".

Catechol- aminergic innervation (norepinephrine and dopamine) of the IO is also not spread evenly in general (Hoffman and Sladek 1973). In the rat, the VLO, known to be involved in compensatory eye movements, is targeted by dopaminergic projections from the mesodiencephalic junction (Toonen et al. 1998). Again, this can differ between species: In the cat, dopaminergic nerve terminals are most prominent in the DAO (Maqbool et al. 1993) whereas catecholaminergic fibers are mostly seen in the MAO and lateral lamella of the PO in the monkey (Sladek and Bowman 1975; Kamei et al. 1981). Interestingly, the noradrenaline fibers in the IO are much more homo- geneously distributed in humans than in rat, cat, and monkey (Powers et al. 1990).

Action potentials of IO neurons have a characteristic waveform (Fig. 43.5b): A sharp spike is followed by a prolonged afterdepolarization (ADP) and a long- lasting afterhyperpolarization (AHP; Llina ́s and Yarom 1981a, b). IO neurons express a low discharge rate. Generally, they spike only once or twice per second (Llina ́s and Yarom 1981a, b; Benardo and Foster 1986). Although the IO neurons are very sensitive to sensory stimulation, their responsiveness is very limited. Stimulating excitatory afferents to the IO either from sensory nuclei such as the trigeminal nucleus or from higher systems such as the mesodiencephalic junction evokes only a single action potential in IO neurons. Using a chemical excitant, such as harmaline, results in a maximal IO firing of approximately 10 Hz (Llina ́s and Volkind 1973; Llina ́s and Yarom 1986). This low discharge rate and limited responsiveness of the IO neurons are unique features in the generally very active olivocerebellar circuit.

The fast IO spike is mediated by Na+, whereas the ADP is generated by the activation of dendritic high-threshold Ca2+ conductances (Llina ́s and Yarom 1981a, b). Recently, Choi et al. (2010) revealed the involvement of the P/Q type Ca2+ channel CaV2.1 in this process. The influx of Ca2+ activates dendritic Ca2+-activated K+ conductances, inducing subsequently the slow AHP (Llina ́s and Yarom 1981a;...

One to seven small wavelets (<10 mV) are superimposed on the ADP of IO neurons (Fig. 43.5b, arrows). The wavelets occur at very high frequencies ranging from 200 to 500 Hz. These small wavelets represent high-frequency bursts of action potentials that are most likely generated in the axon of IO neurons by the ADP (Crill and Kennedy 1967; Crill 1970; Maruta et al. 2007; Mathy et al. 2009). The ADP propagates to the axon where it initiates a burst of action potentials. These action potentials, in turn, back-propagate to the soma. Here, they only give rise to small wavelets, which are superimposed on the ADP. Na+ spikes are not evoked because the Na+ channels are still inactivated at the soma (Mathy et al. 2009). These burst firings of IO axons are transmitted to the Purkinje cells via the climbing fibers, where they can modify the Purkinje cell complex spike itself or the synaptic transmission between the parallel fiber and Purkinje cell (Hansel 2009; Mathy et al. 2009).

Only IO neurons from the DCK and the VLO do not express subthreshold oscillations (Urbano et al. 2006). Frequency modulation of subthreshold oscillation has been observed within IO recordings, but it is more frequently seen in vitro than in vivo (Devor and Yarom 2002b; Khosrovani et al. 2007). The origin of shifts in oscillation frequency is not yet known. However, Placantonakis and Welsh (2001) revealed that activation of NMDA receptors in the IO can support and is able to modulate subthreshold rhythmogenesis in IO neurons. In vivo IO recordings from animals anesthetized by ketamine (blocker of NMDA receptors) express more often LTOs than SSTOs, whereas IO recordings of animals in which the anesthetics used do not block the NMDA receptors (medetomidine/ midazolam/fentanyl mixture) express more often SSTOs than LTOs (De Jeu et al. 2010). Therefore, SSTOs and LTOs might very well be related to the activation of NMDA receptors and sudden switches between SSTOs and LTOs could be induced by an abrupt alteration in glutamatergic inputs. Amplitude modulations have mainly been observed in vitro (Devor and Yarom 2002b). Several studies revealed that a number of neuroactive substances are able to modulate the amplitude of the subthreshold oscillations. For example, serotonin suppresses the amplitude of the subthreshold oscillation (Placantonakis et al. 2000), whereas NMDA and harmaline enlarge the amplitude of the subthreshold oscillation (Llina ́s and Yarom 1986; Placantonakis and Welsh 2001).

Summary: IO as a learning driver is ideal for predictive learning, by having bursting and phasic signaling with long temporal gaps to allow predictions to be expressed.

Subthreshold oscillations determine the firing behavior of IO neurons in
a temporal manner. During the peak of the oscillation, there is an enhanced
probability of spiking, whereas during the trough of the oscillation, there is reduced
probability of spiking. IO neurons do not generate an action potential on every
depolarizing phase of the oscillation; on average the IO neurons spike once every
ten cycles.

The phase of the subthreshold oscillation influences the probability of spiking, but
spiking also influences the phase of the following oscillation (Leznik et al. 2002;
Khosrovani et al. 2007). IO spikes consistently shift the subthreshold oscillation
phase in order to pursuit spiking at the peak of the subthreshold oscillation (i.e., at 90deg)

The electrophys- iological properties of Cx36 gap junctions are characterized by low unitary conduc- tance of 10–15 pS, weak voltage sensitivity, and low pass filter function (Srinivas et al. 1999). Due to these properties, low-frequency membrane oscillations are preferentially transmitted from one IO neuron to another (Fig. 43.5c), resulting in the capacity to synchronize subthreshold oscillations among coupled IO neurons. .. This synchronization of subthreshold oscillations and synchrony of firing among IO neurons is often considered the main function of gap junction coupling.

Gap junctions keep the timing coordinated!

Despite discussion of functional implications and specific discussion of IO as error signal, they somehow fail to recognize that the combination of inhibitory and excitatory inputs to IO is ideal for actually computing such an error signal in the first place!!

LoyolaHooglandHoedemakerEtAl23

Recent paper shows that timing of inhibitory vs. excitatory inputs is key in driving IO behavior: inhibitory input resets intrinsic oscillations and an excitatory spike within ~50ms is effectively cancelled. However, a spike after ~150ms is optimal at driving IO firing.

Hmm. Does this sound like a temporal coincidence detector to anyone else but me? It is exactly what you want to detect an unpredicted outcome (excitation), where the prediction is conveyed by the inhibitory input.

In this context, the job of the cerebellar cortex is, as in Verduzco-FloresOReilly15, to anticipate the error next time around, to drive an exploratory or anticipatory motor action that might fix the error next time around. In this sense, the PC disinhibition is not about precise motor control signals, consistent with the generally crappy signal-to-noise ratio and slow firing rate of PC neurons. All of that happens in the DCN with its higher firing rates and direct inputs etc.

Mesodiencephalic junction (MDJ)

WangNovelloGaoEtAl22

key modern paper showing strong topology in MDJ connectivity

The CN-MDJ-IO loop may well support the CC-MDJ-IO loop in serving these timing functions during bidirectional control. Whereas the MDJ provides a robust excitatory input to the electrotonically coupled dendritic spines of the PO and rostral MAO, the CN pro- vides an equally robust inhibitory input to the very same dendritic spines (De Zeeuw et al., 1989, 1998; Ruigrok & Voogd, 1995). This configuration of a combined excitatory and inhibitory input to excit- able spines within a glomerulus renders the activation of olivary neu- rons particularly sensitive to the timing between these two inputs (De Zeeuw et al., 1998; Negrello et al., 2019). Moreover, as we have shown in the current study, the CN apparently also provides a prom- inent projection to MDJ neurons that is topographically aligned with their input from the CC. Thus, both at the level of cyto-architecture of the olivary neuropil and at the level of the network connectivity of the CN-MDJ-IO and CC-MDJ-IO loops, the climbing fiber system appears well-designed to facilitate precise timing of neuronal activ- ity required for complex functions mediated by cerebro-cerebellar control.

Yes this is called prediction error. dude.

WangLiuAngelovEtAl23

Our finding challenges
the long-standing dogma that the cerebellum inhibits the inferior olivary pathway and provides a new circuit mechanism for the cerebellar control of high-velocity movements.

In contrast, the excitatory output of the intermediate/lateral cere- bellar modules diverges from the inhibitory nucleo-olivary pathway (Fig. 7i, right motif ) and suppresses CS activity once activated. There- fore, the intermediate/lateral cerebellar modules may implement distinct computational mechanisms and operate in different spati- otemporal domains. Indeed, the IO regions that target the intermedi- ate/lateral cerebellar cortices receive prominent excitatory inputs from the mesodiencephalic junction, an extracerebellar hub region that integrates diverse cerebral information48. We postulate that climbing fiber inputs to the intermediate/lateral cerebellar modules convey diverse cerebral information and are likely subject to motor control mediated by the inhibitory nucleo-olivary pathway49.

Within the forward model of motor control framework, the cerebellum is considered to generate predictions for future move- ments1,50,51. Such computation thus requires IO neurons to compare the extracerebellar motor command and cerebellar prediction and to generate CSs when these two mismatches. The error/instructive signals conveyed by climbing fiber inputs are thought to mediate cer- ebellarplasticityusingsupervisedlearning52–54.TheinhibitoryCN–IO projection is particularly useful for preventing saturation or causing extinction27,55. Our study suggests new computational principles for cerebellar learning. The FNE–IO pathway is designed to drive precisely timed feedback CSs, crucial for instructing long-term plasticity in the cerebellar cortical neurons and cerebellar learning55. Given that the FN neurons have relatively high intrinsic firing rates, neurons in the mcMAO are likely to function as high-pass filters that translate specific FN activity patterns into well-timed CSs during learning. Under this sce- nario, cerebellum-driven internal feedback could be directly integrated into the same cerebellar module, or even into the same PC, and induce plasticity within the loop. Such circuit design might be most efficient for providing feedback information regarding the consequence of its computation and thereby optimizing learning via backpropagation47. This form of feedback information may be similar to a prediction error signal during reinforcement learning56,57. Therefore, we hypothesize that the FNE–IO modules are in principle well-equipped to generate internal models, by performing not only supervised learning but also reinforcement learning58–60.

Below we speculate two plausible sce- narios that the FNE–IO modules could contribute to cerebellar learning. First, the FNE–IO modules could contribute to enhancing the medial cerebellar outputs during learning. When the FNE–IO neurons are acti- vated, the increased climbing fiber input to the cerebellar cortex may be considered a form of positive prediction error. Synaptic plasticity in the cerebellar cortex, governed by this positive prediction error, may loop back to the same FN neurons and further strengthen the desired cerebellar outputs during subsequent learning (Supplementary Fig. 11a). Such positive feedback loop is consistent with the gradual emergence of CSs during saccadic adaptation17. Another possibility is that the feedback CSs might carry sensory prediction errors when the expectation computed by the cerebellum does not match with the actual sensory feedback61. Assuming FN functions as a comparator between the cerebellar output and the sensory feedback, it is likely that the mismatch activates FN neurons and subsequently generates feedback CSs via the FNE–IO–PC pathway. Using a rich repertoire of synaptic plasticity rules55, instructed by this CS signal, the cerebellum could finetune the internal model for motor adaptation. Indeed this conjecture is consistent with the finding in monkeys that FN neurons display strong activation after the introduction of perturbation and gradually decline during the adaptive head turning61 (Supplementary Fig. 11b). Interestingly, the excitatory and inhibitory FN–IO modules segregate the cerebellar cortical regions (Fig. 2). Having parallel access to both supervised learning and reinforcement learning may vastly enhance the capacity of cerebellum learning.

YES, finally! (sort of I guess..)

However, the excitatory CN–IO pathway is confined to FN, the phylogenetically oldestCN37,65,presentingacuriouscaseforevolutionaryneurobiology. Indeed, although all three nuclei innervate a large collection of brain regions, the IN and DN projection patterns are more identical than the FN projection patterns34,37. Therefore, the medial cerebello-olivery module possesses a unique circuitry that distinguishes itself from the rest of the cerebellum.

Dude, CN -> MDJ! needed to integrate cortical prediction signals. FN is simple enough to be able to do it directly.

[rcoreilly]

GABA within DCN

UusisaariKnopfel08

This concept has been challenged by hypotheses which suggest the DCN may act as the substrate of motor memory storage (Lis- berger and Sejnowski, 1992; Wada et al., 2007; Ito, 2006). The demonstrations of mechanisms which cause modifi- cation of synaptic strength (Telgkamp and Raman, 2002; Aizenman et al., 1998) and active membrane properties (Aizenman and Linden, 2000) support this latter viewpoint. It has also been proposed that “cerebellar memory” is acquired in the cerebellar cortex and subsequently trans- ferred to the DCN for consolidation (Mauk, 1997; Ohyama et al., 2006; Masuda and Amari, 2008; see also Jirenhed et al., 2007). In addition to these possible roles in cerebellar memory, it appears likely that the neuronal network within the DCN processes sensory and motor information in “real- time” but very little is known about the computational func- tions of the DCN, particularly with respect to the role of specific cell types.

Indeed, morphological and electrophys- iological studies have revealed that the DCN consists of diverse neuronal populations (Chan-Palay, 1977; Chen and Hillman, 1993; Aizenman et al., 2003; Czubayko et al., 2001; Molineux et al., 2006; Uusisaari et al., 2007), with both glutamatergic and GABAergic neurons forming pro- jecting and local connections (Teune et al., 1995).

We have previously defined three types of DCN neu- rons based on their transmitter content and functional properties (Uusisaari et al., 2007). Importantly, all of the cells studied fired action potentials spontaneously in the absence of synaptic input. The large non-GABAergic neu- rons (GADnL; presumed excitatory and glutamatergic projecting neurons) are specialized for conveying tonic spiking activity due to their relative linear and non-adaptive current-frequency relationship. Conversely, the smaller non-GABAergic (GADnS; presumed glutamatergic) and GABAergic (GAD+) neurons exhibit a more pronounced frequency adaptation and lower maximal frequencies dur- ing sustained firing. These appear to be more suited for processing phasic signals (Uusisaari et al., 2007).

At present, it is generally assumed that all DCN cells are uniformly contacted by Purkinje neuron synapses that mainly reside on their soma or proximal dendrites (Chan- Palay, 1977; De Zeeuw and Berrebi, 1995) whereas the glutamatergic afferent input from MFs and climbing fibers (CF) is conveyed via more distal synapses, away from the soma (Chan-Palay, 1977)

Key for plasticity: MF is distal, weaker, and PC is stronger, proximal..

neuron axonal terminals contacting DCN neurons have been histologically demonstrated (Wassef et al., 1986) and there are GABAergic neurons that do not project outside DCN (Chan-Palay, 1977). ... leads us to conclude that these two cases were most likely functional examples of local GABAergic contacts in DCN. Since the local synaptic network would be expected to be relatively intact unlike the cortico-nuclear connection and the GABAergic cells in DCN fire spontaneously under these experimental conditions, the very low occurrence of this kind of large GABAergic events would suggest that the local synaptic connectivity within DCN slice is either very sparse or only very weakly reflected in somatic recordings.

GAD+ inhib get much smaller inhib, unrelated to cell size.

In contrast, the decay tau of the PSCs in GAD� cells measured in the presence of NBQX and APV was up to four times longer than those in GADnL and GADnS cells, and the 10–90% rise times in GAD� cells were almost double compared with GAD� cells (GAD�: decay ��19.01�1.8 ms, n�14 cells; rise time�1.14�0.12 ms, n�11 cells; t-test GAD� vs. GADnS for decay �: P�0.0001; for rise time P�0.0066; solid color bars in Fig. 5B).

We therefore considered the possibility that the slower IPSCs in GAD� neurons originate from GABAARs that are localized on compartments electrotonically remote from the soma. This idea was supported by the different delays after which IPSCs became apparent following formation of whole-cell, from cell attached configuration. While in GADnL and S cells the IPSCs were visible from the very first seconds of whole-cell recording (Fig. 7, upper left and right panels), in GAD� cells they often appeared only after several tens of seconds and kept growing in amplitude and frequency up to 5 min after breaking into whole-cell con- figuration (Fig. 7, lower left). These observations from five GADnL cells, four GADnS cells and nine GAD� cells are summarized in Fig. 7 (lower right panel), with single expo- nentials fitted to the data (� for GADnL, GADnS and GAD� cells 1.6, 1.6 and 18.2 s, respectively).

Taken together, our findings indicate that the GADnS and GADnL cells are the targets of the Purkinje neurons via somatic or proximal dendritic synapses harboring GABAA Rs involving �1 or �1 /�3 subunits, whereas the GAD� cells receive distal dendritic GABAergic inputs via �1/�3-subunit receptors.

The hypothesis that the GAD� cells have GABAergic synapses further away from cell body is not necessarily inconsistent with the work of De Zeeuw and Berrebi (1995), who show that at least some GABAergic cells do have L7-positive and thus Purkinje neuron-originating ter- minals on their cell bodies. Indeed, in addition to the three cell groups examined here and defined in Uusisaari et al. (2007), there are some very small (soma size less than 10 �m) GAD� cells in the DCN, which could be the GABAer- gic neurons projecting to the inferior olive (IO; De Zeeuw and Van Alphen, 1997). The IO-projecting GABAergic cell group has specifically been shown to be targeted by Purkinje neu- rons, whereas the GABAergic cell population discussed here might consist of a distinct group of local interneurons devoid of direct and strong cortical inhibition. Furthermore, it has been shown by Sultan et al. (2002) that the density of GABAergic terminals on GABAergic cells is lower than on non-GABAergic cells. Thus, even though it is possible that there is a cortical projection to the GAD� cells in the DCN, it is significantly weaker than that to the GAD� cells. This suggested difference in the origin of synaptic inhibition be- tween the GABAergic and non-GABAergic cells is reminis- cent of the reported difference in vestibular nuclei (Camp et al., 2006), where the GABAergic type A neurons (Bagnall and Stevens, 2007) receive almost exclusively slow GABAergic inhibitory inputs, whereas the type B neurons are innervated by a mixture of GABAergic and faster glycinergic terminals.

Finally, the very low occurrence of TTX-sensitive post-synaptic events in a structure where all cells are known to be spontaneously active suggests (albeit cir- cumstantially) that unlike in other brain areas such as hippocampus and cerebral cortex the synaptic intercon- nectivity between the DCN cells is either sparse or very weak.

Overall, kind of inconclusive it seems..

KebschullCasoniConsalezEtAl24

wow definitive reference for lots of detailed info about DCN, IO, and c. cortex

however, after all that, it is unclear if there is any significant level of direct GABA inhibition within DCN that could take over what is initially provided by the PC projections, per the "parallel DCN" notion (Herzfeld et al).

Re persistent retention with loss of PC inputs to DCN, this could be similar to lesions of entire cerebellum which depends on learning in the cortex instructed by the DCN signals to the thalamus. Probably safer to just go with that for the time being..

The non-IO-originating glutamatergic afferent inputs are commonly lumped under the label of “mossy fiber CN inputs,” as many (but not all) of the MF axons providing the cerebellar cortex with multimodal and dense representa- tion of the external and internal states of the world, branch and also terminate in the CN (but see “CN Connections of Mossy Fibers”). For brevity, in this section, we refer to the non-IO-originating glutamatergic afferents as “MF”.

Stimulation of brain regions providing MFs, as well as sensory stimulation of various body parts has been shown to increase the spike rate of at least some CN neurons [238–241]. Interestingly, these synaptic connections have been shown to be resistant to classic Heb- bian plasticity-evoking protocols. Instead, MF-CN synapses undergo long-term potentiation only when input bursts are paired with a delayed inhibition-excitation sequence. The time course of this phenomenon could have physiological signifi- cance in a behavioral context in which a transient increase in PC activity, driven by activation of the MF-parallel fiber path- way, would be followed by a pause. This, in turn, would drive an inhibition-rebound sequence in CN neurons [96, 237, 242].

Finally, since evidence has emerged showing terminals of a MF pathway expanding within the CN when the ani- mal undergoes an eyeblink conditioning protocol [234], the MF-CN synapse has gained interest as a possible key site of plasticity involved in cerebellar learning [243, 245–248]. The scarcity of direct experimental investigations of MF-CN plasticity in living animals limits our ability to draw strong conclusions regarding the role of MFs in shaping cerebellar output. However, it appears evident that the rich sensory and motor reality that modulates cerebellar cortical dynamics is also directly available to the CN, underlining the capacity of the CN to encode behavioral trajectories [249–251].

In contrast to the inhibitory nucleocortical projection, the existence of glutamatergic nucleocortical afferents has been much more widely acknowledged [171, 173, 174, 262,265–267]. It is currently unclear whether there are differences in cerebellar cortical targeting between Class- A and Class-B glutamatergic projection neurons, but evi- dence from the mouse Lat suggests that both do project to the cerebellar cortex [35].

DeZeeuwBerrebi95

This is the main citation for distribution of PC -> CN projections, and actually the PC -> CN_GABA_IO projections are relatively very weak compared to what hits the excitatory neurons! Thus the predictor signal is likely NOT affected nearly as much by the same inhibition -- maybe just enough to get something from the complex spike (which also goes directly from IO).

Also the non-PC inhibition is actually on the same scale as the PC input, so I think it is completely reasonable actually on this basis to conclude that local axon collaterals from CN_GABA_IO could drive direct canceling inhibition!

[rcoreilly]

Numbers of inputs to each IO

ZeeuwHoogenraadKoekkoekEtAl98

In general, each olivary subnucleus projects contralaterally to one or more longitudinal zones of Purkinje cells and gives off collaterals to the central cer- ebellar nucleus that receives its main Purkinje-cell input from the same zone (or zones)12–14

So it gets contralateral inputs and the projects back contralaterally, so all that for nothing?

An important aspect of this loop is that the cerebellar nuclei contain both inhibitory and excitatory projection neurons. The inhibitory neurons provide a GABAergic feedback to the inferior olive, while the excitatory neurons are the mediators via which the cerebellum influences motor behavior22–25. One of the major targets of these excitatory neurons is the meso- diencephalic junction. This area incorporates a variety of nuclei, including the nucleus of Darkschewitsch, red nucleus, nucleus interstitialis of Cajal, nucleus of Bechterew, tegmental field of Forel, zona incerta, sub- parafascicularis nucleus, and the prerubral reticular formation25–27. Some of these nuclei, such as the magno- cellular red nucleus, project directly to motoneurons and interneurons in the spinal cord affecting motor activity28, whereas others such as the parvocellular red nucleus, the nucleus of Darkschewitsch, and the nucleus of Bechterew project to the inferior olive29. Recently, it was demonstrated that cerebellar terminals in the nucleus of Darkschewitsch indeed directly innervate those neurons that project to the inferior olive30.

Thus, the olivocerebellar mesodiencephalic loop is formed by the projections from the olivary collaterals to the cer- ebellar nuclei, from the cerebellar nuclei to the meso- diencephalic junction, and from the mesodiencephalic junction back to the inferior olive.

Similar to the organization of the olivocerebellar modules, the olivo- cerebellar mesodiencephalic loop appears to be topo- graphically organized. For example, the rostral medial accessory olive projects to the posterior interposed cer- ebellar nucleus, which in turn innervates the nucleus of Darkschewitsch, which projects to the rostral medial accessory olive, while the principal olive pro- jects to the dentate cerebellar nucleus, which in turn projects to the parvocellular red nucleus and nucleus of Bechterew, which pro- ject to the principal olive.

This potentially reverberating loop could be controlled by local inhibitory interneurons in the cerebellar nuclei and mesodien- cephalic junction, by the GABAergic feedback from the cerebellar nuclei to the inferior olive, and by the Purkinje cell input to the cerebellar nuclei neurons. Importantly, the Purkinje cell input, which is entirely GABAergic31, affects both the excitatory and inhibitory neur- ons in the cerebellar nuclei32,33. In fact, individual Purkinje-cell axons can innervate both the excitatory
neurons in the cerebellar nuclei that project to the mesodiencephalic junction and the inhibitory neur- ons that provide a GABAergic input to the olive. Thus, the Purkinje cells can control simultaneously the exci- tatory reverberating olivocerebellar mesodiencephalic loop and the inhibitory feedback that could also partly control this loop (see also Fig. 2).

In general, a glomerulus contains a core of five or six dendritic and axonal spiny appendages, derived from different neurons, that is surrounded by four or five terminals and several glial sheaths45,49,50. Serial-section analysis has demonstrated that virtually all spines are located in glomeruli (Fig. 3).

Wait, is that 5-6 receiving IO neurons in one glom??

4-5 terminals = inputs.

Although only about half of the ter- minals in the inferior olive are located inside glomeruli, we evaluate the possible functions of the olive from the perspective of the organization of its glomeruli, because the electrical activities generated in the glomerular spines constitute the unique electrophysiological fea- tures of the olivary neuropil.

As such, the olivary spines might function as the substrate of a precise time filter that determines whether or not particular signals are being transmitted at particular moments in time48,57.

YES:

Several studies have demon- strated that olivary cells that are highly responsive to a cutaneous stimulus applied to the paw under resting conditions fail to respond when a cat stimulates its own paw, although the same cells do respond if a stimulus interrupts the trained movement (Refs 14,90,98,99, cf. Ref. 100). This observation indicates that precisely timed mechanisms must be available for modulating the sensory responsiveness of olivary cells. Since, as ex- plained above, the interactions between the excitatory and inhibitory inputs are highly sensitive to the timing between them, the olivary glomeruli might function as the substrate for this filtering process and deter- mine when and when not an input signal can be con- sidered as an error signal, and accordingly transmit it or not.

However, because most studies of the olivocerebellar system have focussed on the long-term changes that occur at the parallel fiber–Purkinje cell synapses3,8,106, or at the synaptic input to the central nuclei neurons5,107–110, evidence for plastic changes associated with learning processes within the inferior olive is fairly limited111,112.

The comparator hypothesis of olivary function, ini- tially advanced by Oscarsson2,113, proposes that the olivocerebellar system compares intended with per- formed movements to provide error detection (for re- view, see Ref. 1). The error-signaling part of this hy- pothesis has experimental support (see above), but with regard to the function of the olivary microcircuitry, the question is where does the comparison resulting in the error signal occur? In the context in which the comparator hypothesis was originally formulated, the inferior olive was a ‘comparator’ of command signals from higher centers with the activities these signals elicited at lower levels in the spinal cord, with the pre- sumption that the descending paths from the motor cortex and midbrain and the ascending paths from the spinal cord converged at the level of the olive113

However, since anatomical tracing studies over recent years have demonstrated that the descending and as- cending projections to the olive generally do not con- verge on the same olivary neurons (for review, see Ref. 45), it appears unlikely that the signals of intention and achievement are directly compared inside the in- ferior olive. From the anatomical point of view, a com- parison inside the olive is much more likely to happen between the excitatory ascending and descending inputs on the one hand and the inhibitory projections derived from the hindbrain on the other hand.

Thus, if the inferior olive itself functions as a comparator, it might do this by comparing the ex- citatory ascending and descending inputs with the inhibitory inputs from the hindbrain, but this concept diverges substantially from the original hypothesis proposed by Oscarsson, and is more in accord with the blocking process proposed in relation to the learning hypothesis described above.

deZeeuw90

phd thesis!

1 IO -> 10 PC

It was found in man, cat and rat that the number of Purkinje cells is about ten times as high as the number of olivary cells (Kreuz- fuchs,'02; Braitenburg and Atwood,'58; Moa- tamed,'66; Armstrong and Schild,'70; Palko- vits et al.,'71; Mlonyeni,'73), and that each Purkinje cell is provided with one climbing fiber (Eccles et al.,'66). One olivo-cerebellar axon fiber, therefore, should give rise to about ten climbing fibers

do the math

• all gap junctions are in glomeruli
• each glomeruli combines about 5-6 IO neurons
• each IO neuron has about 500–1,000 gap junctions, and that two individual IO neurons may be coupled by 10–20 gap junctions
• = 50 glomeruli!

Each IO is detecting different combinations of inputs??

[rcoreilly]

Forward models

Shadmehr20

Complex spikes arise from climbing fibers that originate from cells in the inferior olive (Eccles et al. 1966). The olive cells receive inhibitory projections from neurons in the deep cerebel- lar nuclei (DCN) (de Zeeuw et al. 1988) and excitatory inputs from other regions of the brain (Saint-Cyr and Courville 1982).

AND, mainly, from the DCN itself!!!

neurons in the inferior olive are in the position to compare the desired response that they have received from a noncerebellar region with the actual output of a DCN neuron and produce an error signal that reflects the difference between the two (De Zeeuw et al. 1998; Kitazawa et al. 1998; Simpson et al. 1996; Soetedjo et al. 2008). From a machine learning perspective, this signal is termed a prediction error.

again, it is not the desired response from a noncerebellar region. it is from the DCN and the "mesodiencephalic junction" whatever that is.

First, the DCN neurons receive a rather weak signal from the olive (Lu et al. 2016), a signal that weakens further as the animal reaches adulthood (Najac and Raman 2017). In comparison, the middle layer neu- rons (the P-cells) receive an extraordinarily strong signal from the olive. Second, despite the fact that each P-cell projects to approximately four DCN neurons (in mice) (Person and Raman 2012a), a P-cell receives only a single climbing fiber. This makes it so that a P-cell contributes to potentially erroneous out- puts of four DCN neurons but receives error information from only one olivary neuron.

The weak olivary input to the DCN is puzzling because it suggests that the error signal from the olive will have difficulty acting as a teacher for the output layer neurons. The single climbing fiber to a P-cell is also puzzling because it implies that the error signal is not a fair reflection of the entire error space but rather biased to provide only a limited view.

HerzfeldHallTringidesEtAl20 provides an answer to this: Cereb. Cortical system provides learning signal that then drives learning in the DCN.

Also, McElvainBagnallSakatosEtAl10 shows that learning signals for neurons with high firing rates are different than standard cortical ones -- the P-cell input to DCN is presumably ideal for driving plasticity there. Although some plasticity directly in DCN can apparently happen, it is not ideal..

Thus, when we view the cerebellum from the perspective of machine learning, two puzzles emerge. It is surprising that the error information from the olive is conveyed strongly to the middle layer (the P-cells) but not the output layer (the DCN). Furthermore, it is surprising that the olivary signal to a P-cell consists of a single input that is biased toward a specific part of the error space.

Here, my aim is to use mathematics of machine learning to consider these puzzles and ask why the cerebellum might be organized in this way. I will suggest that the very limited distri- bution of output from a single olive cell to a handful of P-cells organizes the P-cells into populations that share the same teacher (De Zeeuw et al. 2011; Heck et al. 2013): P-cells that learn from the same teacher may project together as a population to a single neuron in the DCN. This organization of P-cells may serve a critical function: it allows the olive not only to convey error information to a small group of P-cells but through synchronous production of complex spikes also to convey the same error information to the DCN neuron that was responsible for the error (Chaumont et al. 2013; Tang et al. 2019).

As a result, each olivary cell is the teacher to a handful of P-cells (Gao et al. 2012), and those P-cells are the teachers for the DCN neuron that produced the erroneous output (Medina and Mauk 1999).

This fits perfectly with the Herzfeld et al model.

Rather than countering plasticity in the P-cells that learned from error, reversal of error will engage learning in a separate population of P-cells, possibly in a different olivocerebellar module. As a result, organizing P- cells into populations that share a common preference for error may be responsible for a paradoxical aspect of behavior: sponta- neous recovery of memory (Criscimagna-Hemminger and Shadmehr 2008; Kojima et al. 2004; Sarwary et al. 2018; Smith et al. 2006).

Indeed, the assumption that climbing fiber activity reflects a prediction error is not universally accepted (Horn et al. 2004) and faces challenges including the fact that the probability of complex spikes poorly encodes the magnitude of prediction error (Catz et al. 2005) and the fact that cerebellar dependent learning can take place without obvious modulation of com- plex spikes (Hewitt et al. 2015; Ke et al. 2009). In some cases, modulation of simple spikes alone is sufficient to pro- duce a form of learning (Nguyen-Vu et al. 2013). Cognizant of these observations, I will develop the theoretical view- point based on the assumption that olivary input provides an error signal and then consider its limitations.

Proceeds to use FF network to understand cerbelleum. This completely negates key point that DCN is fully independent circuit on its own!

Some of the DCN neurons are GABA-ergic and inhibit the olive
(Bazzigaluppi et al. 2012; Lefler et al. 2014). Activity of these
DCN neurons is labeled as aðd�Þ. Another subset is non-GABA- i
ergic and send their axons to the brainstem, superior colliculus,
red nucleus, thalamus, etc. It is with these non-GABA-ergic
neurons that the cerebellum influences behavior.

If the cerebellar output does not match the actually observed collicular activity, then the olivary cell responds, producing spikes in its cerebellar projections (Kojima and Soetedjo 2017, 2018). Because the collicular activity reflects not only the posi- tion of the visual stimulus, but also its reward value (Ikeda and Hikosaka 2003), the difference in the predicted reward value of the stimulus as compared with the observed value will also be an error signal that is conveyed to the cerebellum (Heffley et al. 2018).

Key point that I've overlooked, which is emphasized in Verduzco & O'Reilly, is that sensory vs. motor errors need to be translated appropriately -- DCN doesn't have that ability? different inputs? some stuff about different projections into the interneurons in the cortex??

Thus, given output ai from the GABA-ergic DCN neu-
rons, we assume that the activity in the inferior olive neuron that projects back to the DCN and the rest of the cerebellum (De Zeeuw et al. 1997b) is an error signal that conveys the differ- ence between what was observed (excitatory input to the olive from an extracerebellar region) and what was predicted (inhibi- tory output from the DCN to the olive).

Puzzle P1. Even though non-GABA-ergic DCN neurons are re- sponsible for conveying output of the cerebellum, the error asso- ciated with their activity is not directly a part of the olivary signal back to the cerebellum. How do non-GABA-ergic DCN neurons receive information regarding the error in their activities?

First of all, this is just plain false. but the other way. non-GABA DCN goes to IO too.

This suggests that even when non-GABA-ergic DCN neurons are ðdþÞ
correctly predicting an output ai , if the inhibitory input to the
olive is suppressed, the error signal to the cerebellum returns.

Conjecture C1. A DCN neuron receives feedback from precisely the same olive neurons it inhibits.

Indeed, olivary projections to the DCN are reciprocal: if a
region in the olive projects to a region in the DCN, then that DCN region also projects back to that specific region of the olive (Ruigrok and Voogd 2000).

A different way to consider puzzle P1 is to note that a single P-cell projects to only four to five DCN neurons (Person and Raman 2012b), and critically, this small group of DCN neurons includes both GABA-ergic, as well as non-GABA-ergic neurons (De Zeeuw and Berrebi 1995; Teune et al. 1998).

Conjecture C2. The prediction error conveyed by an olive neu- ron must result in plasticity in at least one olive-projecting GABA-ergic neuron and one non-olive-projecting, non-GABA- ergic neuron. These two DCN neurons should be coupled in the sense that their activities should always be roughly equal.

At this writing, we do not know how the non-GABA-ergic DCN neurons receive their error information. Conjecture C2 provides the possibility that if the activities in the two DCN neu- rons were identical, then the olivary signal that reflects error for one DCN neuron can also be the signal that teaches the other DCN neuron.

The trivial answer is that perhaps DCN neurons have essen- tially static weights and internal biases and do not learn from their errors. However, this is clearly not the case, as evidenced by the work of Mauk and colleagues (Ohyama et al. 2003; Ohyama and Mauk 2001), De Zeeuw and colleagues (Boele et al. 2013; Carulli et al. 2020), and Nagao and colleagues (Shutoh et al. 2006). For example, in a classical conditioning task in which rabbits hear a tone and learn to close their eyes in antici- pation of an aversive stimulus, training produces plasticity in the P-cells (Jirenhed and Hesslow 2016), as well as the DCN, as evidenced by the fact that after conclusion of training, discon- nection of the P-cell input to the DCN (via GABA antagonists) produces the conditioned behavior, albeit at an earlier time with respect to the tone onset (Medina et al. 2001).

Furthermore, this training coincides with mossy fiber axonal growth and synaptic genesis in the DCN (Boele et al. 2013), as well as changes to the extracellular matrix of molecules that surround the synapses that contact DCN neurons (Carulli et al. 2020). Thus experience of error leads to plasticity in the P-cells as well as the DCN. However, given the weak olivary input to the DCN, how is this possible?

Indeed, the synapses that P-cells make upon a DCN cell can undergo plasticity based on the history of the P-cell simple spikes (Telgkamp and Raman 2002). Furthermore, the synapses that mossy fibers make upon a DCN neuron can also undergo plasticity: when a period of excitation (150 ms) in the DCN neu- ron precedes a period of strong inhibition (250 ms), excitatory synapses strengthen (McElvain et al. 2010; Person and Raman 2010; Pugh and Raman 2008). Thus the sequential pattern of ex- citation from the mossy fibers and inhibition from the P-cells can, in principle, serve as a teacher for changing the synaptic weights of the inputs onto DCN neurons (De Zeeuw et al. 2011; Medina and Mauk 2000).

The regression with respect to error generally exhibited two peaks: one at a lead (�223-ms prediction about a future error) and the other at a lag (+227-ms feedback about the past error) (Popa et al. 2017). The r-squares for the delayed response of simple spikes to error were greater than the r- squares for the prediction of future error. The coefficient of the linear fit that predicted the error was often negative of the coeffi- cient that responded to that error.

Thus simple spikes not only can carry error information, but that information is also sufficient to induce adaptation in the DCN, particularly if the simple spikes are synchronized. However, Eqs. 5 and 6 impose a stringent requirement: the net- work must transmit the error that has been made by a single DCN neuron specifically to that DCN neuron. This means that if the P-cells that converge on a DCN neuron carry error informa- tion, then that information must be the difference between the DCN neuron’s output and the desired output. At this writing, it is unclear how the P-cells through their simple spikes might convey this specific error signal.

A complex spike is a significant event at the soma of a P-cell, generating multiple spikelets. However, the P-cell axon trans- mits this event to the DCN via ordinary spikes (Ito and Simpson 1971): typically, one axonal spike at the onset of the complex spike (with near 100% probability) and another one at the final somatic spikelet (�60% probability) (Khaliq and Raman 2005; Monsivais et al. 2005). Thus the information in the climbing fiber is transmitted from a P-cell to the DCN as a sequence of one or two ordinary spikes. This makes it unlikely that a DCN neuron could dissociate between a complex spike and a simple spike.

When the inferior olive is electrically stimulated, at 5- to 10- ms latency there is occasionally a single spike in a DCN neuron (Bengtsson et al. 2011; Hoebeek et al. 2010). This is likely due to the direct excitatory projections from the olive to the DCN. This occasional spike is then followed by suppression of DCN activity that lasts around 50 ms (Fig. 2) (which can be followed by a rebound burst of activity). The suppression is present in many DCN neurons, including output neurons that do not pro- ject to the olive (e.g., non-GABA-ergic neurons that project to the red nucleus) (Hoebeek et al. 2010). If we view olivary stimu- lation as an artificial means to provide error feedback to the cer- ebellum, then that error leaves its impression not via a strong EPSC in the DCN neuron, but surprisingly, via a strong suppression.

A series of experiments by Lang and colleagues (Blenkinsop and Lang 2011; Tang et al. 2016, 2019) provided critical clues regarding how the activity in the olive produced the suppression in the DCN. Blenkinsop and Lang (2011)... In �100 cases, the spike-triggered response suggested that one of the P-cells was connected to one of the DCN neurons. In 70 of the 100 putative P-cell DCN neuron pairs, the occurrence of a complex spike was followed by a suppression in the DCN, a suppression that lasted from 50 to 100 ms (Fig. 3A, right top).

If the error associated with a DCN neuron’s output is con- veyed to it via its parent P-cells, then we should observe com- plex spikes in the parent P-cell following an error by the DCN neuron. There is indirect evidence for this from optogenetic stimulation of P-cells. A 50-ms period of P-cell stimulation results in synchronized production of simple spikes, but when the P-cell stimulation ends, at 80- to 150-ms latency the P-cells produce a complex spike (Chaumont et al. 2013) (Fig. 3B, right top). The interpretation is that when many P-cells are simultane- ously activated, they inhibit their downstream DCN neurons synchronously, preventing these neurons from firing. This sup- pression of the DCN neuron’s activity removes a source of inhi- bition at the olive. The removal of that inhibition allows the olivary cells to reach threshold earlier and fire, resulting in a complex spike in the P-cells (Fig. 3B, right bottom). This inter- pretation is consistent with a closed loop in which P-cells pro- ject to a DCN neuron, which then projects to an olive neuron, which then projects back to the parent P-cell (Fig. 3B), termed an olivocerebellar module (De Zeeuw et al. 1997b).

To test whether unexpected collicular activity was necessary for learning, Kojima and Soetedjo (2018) deactivated a small region of the colliculus using muscimol. Their idea was to elimi- nate the ability of the colliculus to communicate to the inferior olive the presence of an unexpected visual event and thus block the error signal. Notably, they took advantage of the topographic map of the colliculus and focused the disruption on a specific vector in the visual space. They found that the disruption led to near elimination of learning. That is, the animals were unable to learn from that specific error vector.

Finally, to test whether collicular activity is the source of complex spike tuning, Soetedjo et al. (2019) used subthreshold stimulation of the colliculus while the animal was fixating and recorded from P-cells in the oculomotor vermis. They noted that stimulation at a particular location on the colliculus reliably pro- duced a complex spike (at around 15-ms latency). They then recorded the complex spike tuning of each P-cell with respect to visual errors during a saccade task (as in Fig. 8A). Remarkably, when subthreshold stimulation of a particular region of the colli- culus produced complex spikes in a P-cell, the direction of sac- cade that was encoded by that particular region tended to be aligned to the direction of visual error that was preferred by the P-cell.

These results suggest that a major source of complex spikes in the oculomotor vermis is unexpected activity in the superior colliculus, and that CS tuning of P-cells with respect to direction of visual error is likely derived from spatial organization of col- licular neurons with respect to the visual space.

Conjecture C6. A DCN neuron should project to a group of neu- rons that can produce an effect that will remedy the specific error that is of concern to the parent P-cells.

Indeed, in tasks as diverse as arm movements, head movements, and eye movements, a consistent finding has been that error size can change without altering the probability of complex spikes (Ke et al. 2009; Keating and Thach 1995; Soetedjo et al. 2008). From a machine learning perspective, it is quite problematic if the error signal is insensitive to error size.

[rcoreilly]

Plasticity directly in DCN is sufficient, without Cereb. Cortex?

HerzfeldHallTringidesEtAl20

Key paper finally!

• explanation for cascade of learning from PC / Cortex to DCN: PC can do 1 trial quick learning
• but still not getting the predictive error thing in the DCN -- used only to explain data.
• overall, it is all "behavioral" without any direct neural measurements or interventions.. needs a lot more work to substantiate.

Using a circuit-level model of the cerebellum, we link behavioral data to learning’s neural implementation. The four principles are: (1) early, fast, acquisition driven by climbing fiber inputs to the cerebellar cortex, with poor retention; (2) learned responses of Purkinje cells guide transfer of learning from the cerebellar cortex to the deep cerebellar nucleus, with excellent retention; (3) functionally different neural signals are subject to learning in the cerebellar cortex versus the deep cerebellar nuclei; and (4) negative feedback from the cerebellum to the inferior olive reduces the magnitude of the teaching signal in climbing fibers and limits learning. Our circuit-level model, based on these four principles, explains behavioral data obtained by strategically manipulating the signals responsible for acquisition and recall of direction learning in smooth pursuit eye movements across multiple timescales.

In the motor learning paradigm we study, direction learning in smooth pursuit eye movements, neurophysiological data has demonstrated that short-term motor learning, on the order of a single learning trial, occurs at or upstream of Purkinje cells in the floccular complex of the cerebellum (Medina and Lisberger, 2008; Yang and Lisberger, 2013; Yang and Lisberger, 2014). Single-trial learning is tightly linked to climbing fiber inputs, in agreement with the predictions of the classical theory of cerebellar learning (Marr, 1969; Albus, 1971; Ito, 1984). Crucially, Purkinje cells are only two synapses away from the motoneurons that drive the adaptive motor response we measure, sig- nificantly constraining the circuit architecture that implements the transformation from sensory inputs to adapted behavior across short timescales.

However, recent behavioral results have suggested that longer term pursuit direction learning, on the order of hundreds to thousands of trials, may be mediated by multiple sites and/or learning mechanisms (Hall et al., 2018; Yang and Lisberger, 2010). The existence of multiple components in pursuit learning is consistent with behavioral and computational evidence from other cerebellar- dependent motor learning behaviors (Boyden et al., 2004; Ethier et al., 2008; Kojima et al., 2004; Kording et al., 2007; Lee and Schweighofer, 2009; Medina et al., 2000; Smith et al., 2006; Kassardjian et al., 2005; Shutoh et al., 2006). While plasticity is possible across many synapses in the cerebellar circuit (Carey, 2011; D’Angelo and De Zeeuw, 2009; Hansel et al., 2001; McElvain et al., 2010; Mittmann and Ha ̈usser, 2007), a prime candidate for long-term learning is the deep cerebellar nucleus, one synapse downstream of the Purkinje cells. Convergent neurophysi- ological and behavioral evidence across multiple learning paradigms lends credence to role of the deep cerebellar nucleus for long-term storage of cerebellar-dependent memories (Raymond and Medina, 2018; Kassardjian et al., 2005; Shutoh et al., 2006; Lee et al., 2015; Raymond and Lis- berger, 1998; Popa et al., 2016; Broussard and Kassardjian, 2004; Lisberger, 1994). Taken together, these results highlight the idea that multiple neurophysiological mechanisms within the cerebellar circuit may underlie behavioral motor learning across longer timescales.

Doesn't have the internal model error computation in IO -- just a simpler "negative feedback inhibition" kind of thing..

Projections from Purkinje cells in the cerebellar cortex, in turn, inhibit floccular target neurons (FTNs) in the vestibular nucleus (Lisberger et al., 1994). Cru- cially, FTNs are a single synapse from the moto- neurons that drive adapted behavior (Highstein, 1973; Scudder and Fuchs, 1992).

Plasticity appears to be possible at nearly every synapse in the cerebellar circuit (for a review see Carey, 2011),

Carey11 still main ref..

The saturating dependence of single-trial adaptation in our system on instruction magnitude or stimulus strength parallels findings of single-trial learning in other cerebellar-dependent adaptive behaviors (Fine and Thoroughman, 2006; Herzfeld et al., 2014b; Marko et al., 2012; Hutter and Taylor, 2018).

Thus, we think that the fast-learning process for pursuit learning occurs in the floccular complex of the cerebellum (Medina and Lisberger, 2008) and that the floccular target neurons in the vestibular nucleus (Lisberger et al., 1994) are likely the site of the slow-learning process and long-term storage. At this time, we cannot rule out the possible role of additional downstream areas.

There are multiple reasons to think that learning could transfer from the cerebellar cortex to the deep cerebellar nuclei, even though there is currently no direct neurophysiological evidence. Mecha- nisms of plasticity are abundant in the deep cerebellar nuclei and optogenetic modulation of the activity of Purkinje cells can adapt the vestibulo-ocular reflex, presumably through Purkinje-cell induced plasticity in the deep cerebellar nuclei (Jang et al., 2020; Nguyen-Vu et al., 2013). Record- ings from Purkinje cells and deep cerebellar nucleus neurons following long-term vestibulo-ocular gain adaptation lead to the conclusion of neural learning in both sites (Lisberger, 1994) and are entirely compatible with the suggestion that the learned responses in Purkinje cells instruct learning in the deep cerebellar nucleus (Miles and Lisberger, 1981). For both eye movements (Pastor, Pastor et al., 1994) and classical conditioning of the eyelid response (Garcia et al., 1999), inactiva- tion or lesions of the cerebellar cortex prevents adaptation but does not fully abolish learned responses from previous adaptation (McCormick and Thompson, 1984; Perrett et al., 1993), sug- gesting that the residual long-term learning resides outside of the cerebellar cortex. Finally, for learning in the optokinetic response, lesions of the flocculus abolish the learned response completely after short-term learning but only partially after long-term learning (Shutoh et al., 2006).

In the vestibulo- ocular reflex, Kimpo et al., 2005 showed different generalization for stimulus frequency for gain-up versus gain-down learning, suggesting different sites of learning based on modification of different vestibular signals.

The use in the model of different inputs to the sites of plasticity in the floccular complex and at the floccular target neurons suggests two features of the neural sys- tem that run counter to current doctrine. First, the general rule is that Purkinje cells and their targets in the deep cerebellar nuclei receive identical inputs from mossy fibers. In the floccular complex, however, this does not seem to be the case: brainstem areas that send afferent mossy fibers to the ventral paraflocculus region of the floccular complex send only sparse projections to the location of FTNs in the vestibular nucleus (Osanai et al., 1999; Nagao et al., 1997). Second, the general rule in the oculomotor system is the firing rate is linearly related to eye position and velocity, and this is true for mossy fibers in the floccular complex (Miles et al., 1980; Lisberger and Fuchs, 1978). Thus, there is no obvious precedent for neural responses that are unimodally tuned for eye velocity. Lim- ited recordings from FTNs during pursuit learning Joshua et al., 2013 demonstrate a diversity of FTN velocity-dependent responses to a 30 deg/s pursuit stimulus, perhaps consistent with a distribu- tion of preferred speeds across the neural population. We realize that the assumptions in the model may not reflect the actual implementation of the system in the brain, and that the differences in gen- eralization at different sites could reflect properties of synapses or plasticity mechanisms rather than of the input signals that are subject to plasticity.

We have always been troubled by the fact that pursuit learning, like other forms of motor learning (Krakauer et al., 2000; Tseng et al., 2007; Vaswani et al., 2015), never reaches an amplitude that completely eliminates the error created by the imposed instruction. Even after more than 1000 learn- ing trials, the adapted pursuit eye movements shows residual errors than are in excess of 50% of the imposed instruction (Hall et al., 2018).

Via the known double-inhibitory pathway from Purkinje cells to the inferior olive via the deep cerebellar nucleus (Houck and Person, 2014; Najac and Raman, 2015; de Zeeuw et al., 1988), learned depression of Purkinje cell simple-spike firing could reduce the effectiveness of a given movement error by attenuating the probability and/ or duration of climbing fiber inputs to the cerebellar cortex. Previous neurophysiological observa- tions over 100 trials of pursuit learning demonstrate a reduction in the probability of complex spikes, hinting at possible inhibition of the olive (Medina and Lisberger, 2008; Yang and Lisberger, 2017). Our data and modeling render unlikely the alternate explanation that incomplete adaptation is the result of a competition between the learning signals and a restoring force due to forgetting (Smith et al., 2006; Albert et al., 2019). During error-clamp trials, we measured the retention parameter after 1000 trials to be close to 1.0, corresponding to almost complete retention.

We do not understand fully why it is advantageous for the fast-learning process to remain inhibited after learning has been trans- ferred from the cerebellar cortex to FTNs. Perhaps the goal is to force other parts of the pursuit cir- cuit to take over some of the learning under conditions where permanent changes seem desirable, by transferring some of the learning away even from the FTNs. Alternatively, our somewhat rarified stimulus conditions may not be engaging the neural learning circuit fully so that we are seeing only a part of the full learning capability of the system.

ClarkZhangLavond92

• cooling of interpositus (IP) nucleus (relevant DCN for eyeblink conditioning) abolished learning

BrachaZhaoIrwinEtAl01

• gaba vs. picrotoxin in IP had performance + learning effects

ArmengolPorrasRuizEtAl13

Recent studies carried out in our lab show that the tbl/tbl
mouse seems not to have its motor learning capabilities com-
pletely affected (Figure 1). Thus, although slower than wild-type
mice, tbl/tbl mice perform both fall and horizontal bar test suc-
cessfully (Figures 1A,B), learning consistently from the first to
last session. However, tbl/tbl mice were unable to adapt their
motor responses in the vertical pole test, in which they system-
atically failed to learn through the four sessions (Figure 1C).

Carey11

• key review paper on learning in different parts of the cerebellar circuit, which turned up refs below:

WelshYamaguchiZengEtAl05

Systemic delivery of (1R-1-benzo thiophen-5-yl-2[2-diethylamino)- ethoxy] ethanol hydrochloride (T-588) prevented long-term de- pression (LTD) of the parallel fiber (PF)–Purkinje cell (PC) synapse induced by conjunctive climbing fiber and PF stimulation in vivo. However, similar concentrations of T-588 in the brains of behaving mice and rats affected neither motor learning in the rotorod test nor the learning of motor timing during classical conditioning of the eyeblink reflex. Rats given doses of T-588 that prevented PF–PC LTD were as proficient as controls in learning to adapt the timing of their conditioned eyeblink response to a 150- or 350-ms change in the timing of the paradigm. The experiment indicates that PF–PC LTD under control of the climbing fibers is not required for general motor adaptation or the learning of response timing in two common models of motor learning for which the cerebellum has been implicated. Alternative mechanisms for motor timing and possible functions for LTD in protection from excitotoxicity are discussed.

McElvainBagnallSakatosEtAl10

Here, we examine plasticity at a candidate site of motor learning: vestibular nerve synapses onto neurons that mediate reflexive movements. Pairing nerve activity with changes in postsynaptic voltage induced bidirectional synaptic plasticity in vestibular nucleus projection neurons: long-term potentiation relied on calcium-permeable AMPA receptors and postsynaptic hyperpolariza- tion, whereas long-term depression relied on NMDA receptors and postsynaptic depolarization.

Vestibular nerve afferents and postsynaptic medial vestibular nucleus (MVN) neurons fire tonically at high rates in vivo, and synaptic transmission at central vestibular nerve synapses drives remarkably linear increases in postsynaptic firing rates (Bagnall et al., 2008). How might these constraints influence the existence and nature of plasticity at this synapse? The activity patterns required to modify synaptic efficacy in neurons that fire at high baseline rates differ from those that gate plasticity in quiescent circuits (Jorntell and Hansel, 2006; Pugh and Raman, 2009). Synaptic inhibition or membrane hyperpolariza- tion can induce long-lasting potentiation of intrinsic excitability in MVN neurons (Nelson et al., 2003) and can trigger synaptic plasticity in analogous deep cerebellar nucleus neurons (Pugh and Raman, 2006). Therefore, we reasoned that postsynaptic membrane potential might similarly influence vestibular nerve synaptic plasticity.

Our results support the Miles-Lisberger hypothesis of learning in the VOR (Miles and Lisberger, 1981), which proposes that MVN firing responses to head movements could be regu- lated by plasticity of the primary sensory input onto a subset of vestibular nucleus neurons.

Motor learning in the VOR requires the cerebellum, but cere- bellar inactivation experiments indicate that the expression of learned changes resides downstream of the cerebellar Purkinje cells, in the MVN (Kassardjian et al., 2005; Shutoh et al., 2006).

Another key ref above ^ -- MVN is where it happens, even if learning is facilitated by Purkinje.

Recordings from vestibular nucleus neurons, together with behavioral analyses of motor learning, indicate that parallel modifiable and unmodifiable pathways mediate performance and adaptation of the VOR (Lisberger, 1984; Lisberger and Pa- velko, 1986). The modifiable pathway includes neurons that are powerfully inhibited by cerebellar Purkinje cells and project to ocular motoneurons, and the faster, unmodifiable pathway is thought to include neurons that do not receive cerebellar inhibi- tion but also project to ocular motoneurons (Lisberger and Pa- velko, 1988; Ramachandran and Lisberger, 2008; Scudder and Fuchs, 1992). Increases or decreases in the gain of the VOR are associated with bidirectional changes in the head move- ment-driven firing responses of MVN neurons that receive cere- bellar inhibition. Our findings that bidirectional plasticity of vestibular nerve synapses can only be evoked in a distinct subset of projection neurons are consistent with the notion of modifiable and unmodifiable VOR pathways.

These results indi- cate that fluorescently labeled MVN neurons in the YFP-16 line are predominantly projection neurons, whereas those in the GIN line provide local inhibition within the bilateral MVN.

The 100 Hz stim protocol evoked postsyn- aptic firing, which averaged 38 ± 10 Hz, and induced a short- lasting posttetanic depression, followed by a robust LTD of the vestibular nerve synapse that reduced the peak EPSC amplitude to 0.82 ± 0.06 of the baseline 15 min after protocol (Figure 2A; n = 11, p = 0.03 versus no-protocol control recordings: 0.98 ± 0.01, n = 9). Thus, the vestibular nerve synapse displays activity-dependent, long-term strength modifications, and coin- cident presynaptic and postsynaptic activity results in LTD.

Flat 100hz = LTD

Remarkably, hyperpolarization of postsyn- aptic YFP-16 neurons during high-frequency vestibular nerve stimulation reversed the direction of long-term plasticity and resulted in LTP, which increased the peak EPSC amplitude to 1.31 ± 0.13 of the baseline (Figure 2B; n = 21, p = 0.05 versus no-protocol control; average hyperpolarization: �18.2 ± 1.3 mV).

100hz + inhib = LTP (positive temporal derivative)

The NMDA-R current duration, 104 ± 18 ms, and voltage dependence (Figures 5A and 5C) suggested that NMDA-Rs predominantly contain the NR2A subunit (Cull-Candy and Leszkiewicz, 2004). These data show that two complementary Ca2+-permeable glutamate receptors support vestibular nerve transmission: CP-AMPA-Rs, which pass Ca2+ maximally at rela- tively hyperpolarized postsynaptic membrane potentials; and NMDA-Rs, which pass Ca2+ maximally at relatively depolarized potentials.

Indeed, both LTD and LTP were blocked by the inclu- sion of the Ca2+ chelator BAPTA (5 mM) in the recording pipette (100 Hz stim protocol: 1.00 ± 0.03, n = 7, p = 0.03; 100 Hz stim plus hyperpolarization protocol: 0.98 ± 0.03, n = 6, p = 0.02; Figures 5F and 5I).

Thus, postsynaptic NMDA-Rs are required for LTD, but not LTP, at the vestibular nerve synapse.

Interestingly, the 100 Hz stim plus hyperpolar- ization protocol resulted in synaptic depression (0.81 ± 0.10, n = 5, p = 0.008; Figure 5K) in the presence of PhTx-433, rather than potentiation, indicating that CP-AMPA-Rs are required for LTP induction.

This protocol induced LTP of 1.27 ± 0.12 (Figure 6A; n = 7, p = 0.05), comparable to LTP induced by hyperpolarization lasting 250 ms (p = 0.81). We then extended the hyperpolarization step duration to 1550 ms so that rebound followed synaptic stimulation by 1000 ms. This protocol, which temporally dissociated synaptic input and intrinsic rebound, did not induce LTP (Figure 6B; 0.96 ± 0.09 of baseline, n = 7, p = 0.61 versus control). These results suggest that intrinsic rebound is a critical component of LTP induction that must be temporally linked to synaptic stimulation.

These results indicate that, as in projection neurons, CP-AMPA-R activation paired with postsynaptic hyperpolariza- tion can induce LTP in interneurons, but LTP is typically masked by NMDA-R-mediated LTD.

temporal derivative again:

Although LTD can be induced by nerve afferent activity alone, the induction of LTP requires a specific temporal conjunction of vestibular nerve synaptic activity and postsynaptic hyperpolarization

PayneFrenchGuoEtAl19

Random ref here: Purkinje cells can have pretty high firing rates!

NelsonKrispelSekirnjakEtAl03

Vestibular nucleus neurons fire spon- taneously in intact animals at rates of 50–100 spikes/s (Fuchs and Kimm, 1975) and signal information by syn- aptically induced modulations in firing both above and below these baseline rates.

Synaptic inhibition from the contralat- eral vestibular nucleus is also thought to play a role in VOR plasticity (Darlington et al., 2002).

PughRaman06

Definitely lose timing info on cerebellar lesion in DEC:

Plasticity is not lim- ited to Purkinje cells, however; when the cerebellar cor- tex is disrupted after acquisition of conditioned re- sponses, conditioned stimuli still evoke ill-timed yet stimulus-specific conditioned responses (Garcia and Mauk, 1998; Garcia et al., 1999; Lavond, 2002; Ohyama et al., 2003; Bao et al., 2002). These ‘‘short-latency re- sponses’’ suggest that the storage of memories also in- volves synaptic changes within the cerebellar nuclei.

Hypotheses about the nature of cellular plasticity in the cerebellar nu- clei, however, have been generated with a large-scale model of cerebellar function, which incorporates the known intrinsic and synaptic properties of cerebellar neurons (Medina and Mauk, 1999; Medina et al., 2001, 2002b). The model predicts that plasticity at mossy fi- ber-to-nuclear cell synapses is unlikely to result from correlated pre- and postsynaptic activity. Instead, inhi- bition from Purkinje cells may control potentiation of mossy fiber synapses (Medina and Mauk, 1999), much as it is proposed to regulate excitatory synaptic strength in the vestibular nuclei during plasticity of the vestibulo- ocular reflex (Miles and Lisberger, 1981). Consistent with this idea, disruption of the cerebellar cortex prevents acquisition of new conditioned responses, although pre- viously acquired short-latency responses may persist (Garcia et al., 1999; Bao et al., 2002).

Inhibition from PC = learning signal here, but could be also in direct..

note this is not GABA -> IO

The results indicate that EPSCs are potentiated only when high-frequency syn- aptic activation is coupled with sufficient postsynaptic hyperpolarization to evoke a postinhibitory rebound cur- rent, even in the absence of postsynaptic spiking.

Mossy fibers fire rapidly, with reported rates ranging from 75 to 200 spikes per second (Van Kan et al., 1993; Chadderton et al., 2004),

The first neuronal change that occurs during training ap- pears to be an increase in Purkinje cell firing rate early in the tone and a decrease late in the tone (Hesslow, 1994; Medina and Mauk, 1999; D. Jirenhed et al., 2005, Soc. Neurosci., abstract). Therefore, at the level of the nu- clear cells, which receive GABAergic input from Purkinje cells, this initial modification of activity is predicted to produce concurrent excitation (via mossy fibers) and in- hibition (via Purkinje cells) early in the tone, followed by excitation and primarily disinhibition toward the end of the tone (Figure 1A). Repeated trials of this pattern of ac- tivity drive plasticity of nuclear synapses (Medina and Mauk, 1999).

Cerebellar nu- clear neurons were silenced during the hyperpolariza- tion and responded to the relief of hyperpolarization with a postinhibitory rebound burst of action potentials (Aizenman and Linden, 1999; McKay et al., 2005), firing an average of 36 6 6 spikes while the mossy fiber stimu- lation continued (Figure 1B, bottom, n = 13).

After 30 rep- lications of the induction protocol, which we called the ‘‘MF-rebound’’ protocol, EPSC amplitudes often tran- siently decreased relative to control. Over the next few minutes, EPSCs increased, generally reaching a stable value after 5–7 min, and the potentiation persisted for the duration of the recording (R20 min, Figure 1C).

after the 5–7 min recovery period, increased by 43% 6 13% (n = 13, p = 0.008, Figure 1D). EPSCs were potentiated robustly (by R15%) in 10 of 13 cells, indicat- ing that the induction protocol was reliable.

Consistent with a requirement for ele- vation of internal Ca2+ for induction of potentiation, a pi- pette solution that included 10 mM BAPTA prevented the MF-rebound stimulation from inducing potentiation of EPSC amplitudes (Figure 2A). In fact, on average, EPSCs depressed, decreasing by 22% 6 10% (Fig- ure 2C, n = 5, p = 0.04). When the MF-rebound stimula- tion was applied in the presence of the NMDA receptor antagonist CPP (10 mM, with normal intracellular calcium buffering), EPSCs failed to potentiate, remaining at 93% 6 8% of control values (n = 11, p = 0.71; Figures 2B and 2C). These results suggest that NMDA receptor activation provides a necessary source of the Ca2+ re- quired for potentiation.

We therefore repeated the experiments with an ‘‘MF-only’’ induction protocol com- posed of the 550 ms mossy fiber stimulation with no hy- perpolarizing step; a similar pattern of stimulation has been shown to increase intrinsic excitability in these neurons (Aizenman and Linden, 2000). With this proto- col, cerebellar nuclear neurons fired strongly during each tetanizing stimulus (68 6 13 spikes, Figure 3A, top). Consistent with previous reports (Aizenman and Linden, 2000), however, EPSCs remained at 110% 6 6% of control values (n = 7, p = 0.27; Figure 3A, bottom, and Figure 3C). Thus, potentiation of these synapses does not appear to follow a simple Hebbian rule, in which any high-frequency mossy fiber stimulation that reliably elicits spiking would be expected to potentiate EPSCs. Instead, the hyperpolarization and/or the subse- quent depolarization in the MF-rebound protocol appar- ently contribute(s) a necessary component to the induc- tion of plasticity.

This is all strongly indicative of temporal derivative learning!

This ‘‘rebound- only’’ stimulation produced a small but significant in- crease of 11% 6 5% in EPSC amplitudes (n = 7, p = 0.03; Figure 3B, bottom, and Figure 3C), suggesting that Ca2+ through voltage-gated channels can influence mossy fiber synaptic strength.

When the 250 ms hyperpolarization preceded a 300 ms mossy fiber stimulation such that synaptic excitation overlapped with the rebound depolarization, EPSCs con- sistently failed to potentiate. Instead they remained at 92% 6 4% of control (Figure 4A, middle and right, n = 5, p = 0.11), much like the results obtained by stimulating the mossy fibers with no associated hyperpolarization.

These re- sults provide evidence against the idea that Ca2+ influx during the rebound burst sums with the synaptic Ca2+ in- flux to bring the Ca2+ concentration past a threshold level that triggers plasticity. Instead, they suggest that the specific spatial or temporal profile of Ca2+ influx obtained with the MF-rebound protocol is particularly appropriate for inducing potentiation.

Yep!

we tested the induction protocol in which mossy fiber stimulation was interrup- ted by a 150 ms hyperpolarization without synaptic exci- tation such that stimulation lasted for 250 ms preceding and 250 ms following the step. This pattern of stimula- tion was highly effective, with EPSCs increasing by 62% 6 23%, and seven of eight cells potentiating by R15%

it seemed possible that an induction protocol in which the 250 ms mossy fiber stimulation was restricted to the 250 ms hyperpolarization might also be capable of potentiating EPSCs. Indeed, with this protocol, EPSCs increased by 30% 6 14%, with R15% potentiation oc- curring in 5/11 cells

The most striking aspect of this protocol, however, is that it demonstrates that potentiation can occur even when synaptic excitation does not coincide with post- synaptic spiking. This result, along with the calcium de- pendence of potentiation, raises the possibility that syn- aptic currents at negative potentials contribute to the Ca2+ influx that triggers plasticity. Although in many cen- tral neurons NMDA receptors are strongly blocked by Mg2+ at negative potentials (Mayer et al., 1984, Nowak et al., 1984), cerebellar nuclear cells express the NR2D subunit, which is less sensitive to Mg2+ block (Monyer et al., 1994; Akazawa et al., 1994; Momiyama et al., 1996; Kuner and Schoepfer, 1996; Cull-Candy et al., 1998). We therefore examined the characteristics of syn- aptic currents at hyperpolarized potentials by evoking EPSCs either with single stimuli or with trains.

This result suggests that a substantial synaptic Ca2+ influx may occur during the hyperpolarization itself, especially during high-fre- quency stimulation. Although the amount of Mg2+ block is sufficient to make the Ca2+ influx lower at hyperpolar- ized than at moderately depolarized potentials (see Dis- cussion), the substantial activation of NMDA receptors at negative potentials raises the possibility that postsyn- aptic spiking may not be as central to the induction of plasticity in nuclear cells as it is at many other excitatory synapses.

To test this idea, we designed a voltage-clamp induction protocol in which nuclear cells were held at 265 mV be- fore being stepped to 280 mV for 150 ms and then back to 255 mV for 750 ms. Mossy fiber stimulation began 200 ms before the hyperpolarization and continued for 150 ms following the step to 255 mV (Figure 6B, top panel). With this protocol, EPSCs increased by 30% 6 13%

To test this idea, we designed a voltage-clamp induction protocol in which nuclear cells were held at 265 mV be- fore being stepped to 280 mV for 150 ms and then back to 255 mV for 750 ms. Mossy fiber stimulation began 200 ms before the hyperpolarization and continued for 150 ms following the step to 255 mV (Figure 6B, top panel). With this protocol, EPSCs increased by 30% 6 13%

These results rule out a direct role of hyperpolarization and instead support the idea that the rapid activation of rebound current, likely including low-voltage-acti-vated calcium channels, is necessary to induce robust potentiation of mossy fiber EPSCs in nuclear cells.

it seems likely that plas- ticity is triggered by synaptic calcium influx through NMDA receptors followed by nonsynaptic calcium influx through low-voltage-activated calcium channels.

In contrast, NMDA receptors that are only weakly sensitive to Mg2+, like those in the cerebellar nuclei, would be expected to have a reduced ability for coincidence detection, as they respond to glutamate even at negative potentials.

instead, they fire action potentials spontane- ously, even without synaptic excitation (Thach, 1968; Jahnsen, 1986), largely because of a tonic cation con- ductance that constantly drives the membrane potential above threshold (Raman et al., 2000).

Such disinhibition could result either from learning-associ- ated changes in synaptic drive onto Purkinje cells (Me- dina and Mauk, 1999) or possibly directly from the com- plex spike-induced pause, both of which depend on climbing fiber activity.

Similarly, at parallel fiber- to-Purkinje cell synapses, long-term depression is driven by large Ca2+ influxes, while long-term potentia- tion is induced by small ones (Coesmans et al., 2004; see also Hansel et al., 1997). Our data suggest that the temporal pattern of Ca2+ influx is a relevant variable; the most effective signal for potentiation may be an ini- tial ‘‘priming’’ Ca2+ influx primarily through synaptic re- ceptors followed by a distinct ‘‘triggering’’ Ca2+ influx recruiting voltage-gated calcium channels during the rebound.

This is exactly the temporal derivative!

In these cells, modulating tonic Ca2+ levels by de- creasing firing rates produces long-lasting changes in intrinsic excitability, a form of plasticity that may under- lie adaptation of the vestibulo-ocular reflex (Nelson et al., 2003, 2005).

PughRaman08

followup to 06

Each induction protocol consisted of a 250 ms, 133 Hz stimulation of excitatory afferents (the “synaptic” stimulus) and a 150 ms hyperpolarization to approximately �85 mV, which was followed by spontaneous, postinhibitory action potentials (the “rebound”). The onset time of the synaptic stimuli was var- ied from 600 ms before the rebound to 150 ms after the rebound.

ZhangLinden06

Similar to above: just stimulation of inputs drives LTD. Is mGluR1 dependent.

PughRaman09

great review paper on all the physiological properties of CN neurons.

Recordings from these neuronal somata, which continue to fire spontaneously in isolation, demonstrate that the cells have few K+ channels open at subthreshold potentials. Instead, the primary intrinsic conductance that is active at hyperpolarized voltages is a small, voltage-independent, non-specific cationic leak with a reversal potential near �30 mV, which is referred to as the ‘tonic cation current’ or ‘undercurrent’ [9]. The molecular identity of the channel responsible for this current is not known, but it shares many properties with the recently identified channel NACLN1 [19].

cerebellar nuclear cells simply try to rest above threshold.

In other words, cerebellar nuclear neurons often need to be hyper- polarized to threshold; under physiological conditions, it is likely to be the tonic inhibition from Purkinje cells that keeps them firing.

As a result, upon fusion of a single vesicle, GABA reaches not only the postsynaptic receptors immediately opposite the release site but also receptors in neighboring postsynaptic densities, before being taken up by astrocytic transporters. Because of this ‘spillover’ of neurotransmitter, even a very low release probability of each presynaptic vesicle generates a high probability that GABAA receptors at most postsynaptic densities opposite a single bouton will be exposed to GABA on the time scale of the rising phase of the IPSC [26,27].

Because most T-type currents remain largely inactivated at vol- tages more depolarized than �70 mV [41], the contribution of T-type currents to post-inhibitory rebound responses, at least as recruited by fast synaptic inhibition, might be small.

Consistent with this idea, recent work indicates that stimulation of Purkinje afferents in vivo does not always activate post-inhibitory bursts of action potentials [40]. Nevertheless, the high incidence of T-type Ca2+ channels targeted to the dendrites of nuclear cells indicates that physiological conditions that favor their activation must exist.

Olivary cells fire spontaneously at �1–10 Hz [45]; their axons form climbing fibers, which form synapses onto Purkinje cells, the activity of which leads to long-term depression (LTD) of parallel fibers [46].

TODO: background of spontaneous IO is key for plasticity?

Behavioral and modeling studies indicate that this learning process occurs in two stages, with synaptic and/ or excitability changes occurring first at the level of Pur- kinje cells [52,53] and second at the level of nuclear cells [54–56]. The nature of the long-term plasticity in the cerebellar nuclei (or indeed whether it occurs at all) has been a matter of debate [57,58]. One proposal with con- siderable experimental support from whole-animal studies is that learning involves an instructive signal from Pur- kinje cells that ultimately strengthens mossy fiber inputs onto nuclear neurons [43].

MedinaMauk99

When plasticity at the mossy fiber synapse is controlled by nucleus or climbing fiber activity, the circuit is unable to retain memories because of interactions within the network that produce spontaneous drift of synaptic strength. In contrast, a plasticity rule controlled by the activity of the Purkinje cell allows for a memory trace that is resistant to ongoing activity in the circuit.

good point: any purely self-driven mechanism will be intrinsically unstable.

Key limitation of model: does not include coincidence detection property of IO glomeruli! so CN -> IO is a direct influence vs. being critically dependent on excitatory input.

[rcoreilly]

Adaptive filter model

Ok, after all that, it looks like the basic adaptive filter model is really the story all along. Maybe there is some further value in articulating some of the details which it seems are not yet represented in the literature I've seen so far..

Terminology:

• Forward model / internal model: any predictive learning system
• Adaptive filter: specific use of predictive learning / forward model to cancel sensory reafference (self-motion cancellation).

The inhibitory relationship of PC -o DCN specifically supports the adaptive filter version, as do the evolutionary analogues of cerebellum and the initial vermis / vestibular function. Reflexes are a kind of side-effect.

Adaptive filter also has the advantage of a precisely specified error signal and no issues of translating motor / sensory coordinates: sensory error of any sort can drive generalized error signal that learns to fix -- todo: need to verify this!

Fujita82

Very clear explicit mathematical model, applied to VOR, but maybe too much math..

KawatoGomi92

Thus, in our model, climbing fiber re- sponses are considered to convey motor errors in the motor-command coordinates rather than in the sensory coordinates.

This is a key divergence from adaptive filter.

Kawato99 is more complex combination of inverse and forward, with learning specifically on motor signals.

MiallWeirWolpertEtAl93

• Smith predictor (see fig below)
• Has extensive history of prior theories

MiallWolpert96

Early, widely-cited paper on forward models, includes adaptive filtering / reafference canceling.

Also talks about JordanRumelhart92 forward models to train motor actions in sensory space, and to accomplish desired outcomes in sensory space. "Inverting the forward model" has now been replaced with the Rubicon goal-driven model in my approach..

Bastian06

Nice review of experimental data on prediction errors with lesion patients and ambiguity of motor vs. sensory predictions.

Ito08

• general review paper

ZobeiriCullen24

First definitive evidence of internal model / adaptive filter functionality outside of fish? This is kinda shocking, coming from the world's expert in this phenomenon:

However, the source of the sup- pression signal that cancels peripheral vestibular input to central pathways during expected self-motion remains unknown. Notably, given that vestibular nuclei neurons can explicitly distinguish expected from unexpected vestibular stimuli even when experienced simultaneously11–14, there is no evidence for presynaptic inhibition of the afferent central neuron synapse (reviewed by ref. 15). Taken together, then, these findings raise the question: what is the central source of the sensory suppression signal required to cancel peripheral vestibular input and how is it computed?

Anterior vermis Purkinje cells send strong direct inhibitory inputs to both the vestibular and deep cerebellar nuclei25,26. Furthermore, in response to unexpected self-motion, these Purkinje cells integrate vestibular and proprioceptive stimuli27,28 to mediate the transformation of vestibular information into the body- centric reference frame required for accurate postural control28. To date, however, it is currently unknown how these neurons respond when self-motion is actively generated and thus expected.

Second, using a simple population-based model, we then demonstrate that the convergence of the signal transmitted by ~40 Purkinje cells accounts for the suppression of self- generated vestibular signals in their target neurons in the vestibular and deep cerebellar nuclei. Taken together our results constitute the direct demonstration of the widely held view that the cerebellum integrates movement-based predictions with actual sensory feedback to cancel the sensory consequences of active self-motion.

Our findings above establish that the head motion response of Purkinje cells in the anterior vermis are markedly attenuated in the active condition.

This suggests that actually there is learning in the DCN that doesn't need PC input for suppression.. need to dig further.

To test this proposal, we first tested if we could predict a given neuron’s sensitivity to active head-on-body motion (Fig. 5a, right column) based on the linear summation of its responses to comparable passive head motion (Fig. 5a, left column) and its response to neck motor commands alone measured during attempted head movement (Fig. 5a, middle column). As illustrated in Fig. 5a, a simple linear model well predicted the response of two representative Purkinje cell’s response during active head motion.

At first glance, these find- ings are surprising as they show that an individual Purkinje cell is not able to provide the cancellation signal, given that Purkinje cells send inhibitory projections to the deep cerebellar and vestibular nuclei25,26, which also show attenuated responses to active head motion (reviewed in ref. 30).

Interesting..

contrast to the Purkinje cells of our present study, data for all rFN neurons (100%) are constrained to the yellow shaded region (Fig. 6b, center). The superposition of data from both populations (Fig. 6b, right) emphasizes that the attenuation of Purkinje cells is hetero- geneous compared to their target neurons in rFN. Thus, taken together these findings are consistent with our hypothesis that, because Pur- kinje cell responses are not only attenuated but also spans both directions during active as compared to passive motion, they can generate the cancellation signal required to suppress vestibular reaf- ference in their target neurons (Fig. 6a, Hypothesis 3).

Accordingly, we next tested whether this was the case. Specifi- cally, to generate an inhibitory output that could (i) explain the sup- pression of vestibular input observed in target neurons during active motion (Fig. 7) and also (ii) eliminate the direct influence of motor commands on downstream targets, we used a simple linear model optimizing the weights of the activities of multiple Purkinje cells (Supplementary Fig. 9, see Methods). The latter condition is an important constraint, since target neurons in the vestibular and deep cerebellar nuclei - unlike Purkinje cells - are insensitive to motor commands6,13.

As noted above, Purkinje cells demonstrate considerable hetero- geneity in their response dynamics to both vestibular and proprio- ceptive stimulation28. For example, Purkinje cells can i) show linear, V-shaped, or rectified tuning to vestibular input, ii) be either sensitive or insensitive to passive proprioceptive stimulation (i.e., bimodal vs. unimodal Purkinje cells), or iii) demonstrate excitatory versus inhibi- tory responses to ipsilaterally directed rotational vestibular stimula- tion (type I vs. type II responses).

Finally, in addition to their Purkinje cell inputs, the fastigial nucleus also receives mossy fiber input from the vestibular nuclei, reticular formation, and central cervical nucleus - areas that encode both vestibular and neck proprioceptive sensory information32–35. Accordingly, we tested whether accounting for this additional input altered our Purkinje cell population modeling results. Notably, the dynamics of these inputs have not been characterized during the head and/or neck rotations applied in the present study. Accordingly, we simulated mossy fiber input as a summation of vestibular and neck proprioceptive inputs where the gains and phases were randomly drawn from a distribution comparable to that previously reported in the vestibular nuclei36, see Methods). We then further explored the effect of systematically altering this simulated mossy fiber input rela- tive to the reference distribution of mossy fiber inputs by (i) doubling the gain, (ii) reducing the gain by half, (iii) doubling the phase, and iv) reducing the phase by half (see ref. 28). Overall, we found that the addition of such simulated mossy fiber inputs did not dramatically alter our estimate of the Purkinje cell population size required to suppress the vestibular reafference in rFN neurons during voluntary movements (~50 versus 40; Supplementary Fig. 13). Likewise, com- parable results were obtained for model weight distributions as shown above (Supplementary Fig. 14).

On the one hand, the motor commands generated by the vestibulo-spinal reflex path- ways are counterproductive when the behavioral goal is to actively move through space. Conversely however, the motor commands generated by the vestibulo-ocular reflex (VOR) are vital for effectively stabilizing gaze relative to space during our everyday activities. Indeed, consistent with the functional goals of these pathways, single- unit studies have established that efficacy of the VOR pathway is intact during active head turns when gaze is stable44,45, while that of the vestibulo-spinal reflex pathways is markedly suppressed6–11. It is thus essential that vestibular afferents transmit a robust and accurate representation of head motion to central VOR pathways to ensure gaze stability regardless of whether head motion is passively or actively generated, as has been previously demonstrated1–4. In this context, our present results provide direct evidence that vestibular reafferent suppression is mediated centrally rather than peripherally. Specifically, we show that the output of the cerebellar anterior vermis is consistent with a forward model that selectively suppresses the reafferent com- ponent of the signal transmitted by vestibular afferents to the vestib- ular and deep cerebellar nuclei.

Impor- tantly, the presence of these motor-related signals in the anterior vermis contrasts with their complete absence at the level of their target neurons in both the vestibular and deep cerebellar nuclei11,13. Prior anatomical studies have established that cortical areas including pri- mary motor cortex (M1) project to the anterior vermis51. Indeed, pro- jections from motor cortex inputs contribute to the suppression of sound-evoked responses in the mouse auditory system during movement52,53. Additionally, movement signals could originate in subcortical structures, for example the lateral reticular nucleus54, higher-order thalamus55, and basal ganglia56, since these areas encode movement as well as sensory information. Irrespective of the source of the motor signals, the presence of motor-related inputs in these Pur- kinje cells is consistent with a theorized role of the anterior vermis in generating an internal forward model of self-motion16,57–59.

Interesting, our finding that the firing rates of deep cerebellar nuclei neurons can be reconstructed as a weighted linear sum of cer- ebellar Purkinje cell (and Mossy fiber) inputs aligns with Tanaka et al.‘s research65,66 on wrist movements. In their study, they were also able to predict the responses of Mossy fibers based on the activity of deep cerebellar nuclei from the previous trial, in a manner similar to a Kal- man filter. Given that the suppression of vestibular reafference in our system represents direct evidence of a computation reliant on the internal forward model,

how these synaptic weights change during a learning task (i.e. ref. 67,) is interesting direction for future work.

First, the encoding of sensorimotor inputs is advantageous in the context of motor learning, since it provides redundancy in the synaptic weights that generate a new desired outcome (i.e., learning in weight space, reviewed in ref. 68). When adaptation to a new environment is required, fast and reliable learning can be achieved by small chan- ges distributed across individual Purkinje cells.

MontgomeryPerks19

Peripheral sensory receptors often do not distinguish externally generated stimuli from stimuli generated by an animal’s own behavior (termed reafference). This problem of reafferent noise is one that many sensory systems face. For example, electroreceptor afferents in elasmobranch fish that are specialized for detecting low-frequency bioelectric fields such as those generated by prey are also strongly modulated by the animal’s breathing. However, the output of principal neurons in the DON (so-called ascending efferent neurons; AENs) show essentially no response to the animal’s breathing, demonstrating that a solution to the problem exists (see Figure 1). The detailed mechanism of this solution depends on two steps: (a) subtraction of common input across receptors and (b) generation of a cancellation signal specific to predictable input.

A significant component of reafference is distributed across many receptors and their afferents, which enables “common-mode subtraction” (Bodznick & Montgomery, 1992). Common-mode subtraction is the principle used in differential amplifiers. When recording a signal in the presence of noise, a separate electrode is used to detect the noise alone that is then subtracted by the differential amplifier. This common-mode subtraction is imple- mented in each AEN by the combination of two distinct sources of input: (a) a few afferent fibers corresponding to a spatially discrete excitatory receptive field that convey both external signal and self-generated noise and (b) inhibitory interneurons that pool a large number of afferent input such that they preferentially respond to self-generated noise common across these inputs rather than to spatially discrete external signals driving only a few. Spatially broad inhibition then subtracts spatially distributed self-generated noise from spatially restricted afferent input at the level of each AEN (see Figure 1).

TanakaIshikawaLeeEtAl20

Three lines of supporting evidence for the internal-forward- model hypothesis are presented in detail. First, we provide physiological evidence that the cerebellar outputs (activities of dentate nucleus cells) are predictive for the cerebellar inputs [activities of mossy fibers (MFs)]. Second, we provide behavioral evidence that a component of movement kinematics is predictive for target motion in control subjects but lags behind a target motion in patients with cerebellar ataxia. Third, we provide morphological evidence that the cerebellar cortex and the dentate nucleus receive separate MF projections, a prerequisite for optimal estimation. Finally, we speculate that the predictive computation in the cerebro-cerebellum could be deployed to not only motor control but also to non-motor, cognitive functions. This review concludes that the predictive computation of the internal forward model is the unifying algorithmic principle for understanding diverse functions played by the cerebro-cerebellum.

Nice additional motivation:

The cerebellum has developed in the sensory domain of the central nervous system (CNS) as evidenced by the fact that it has emerged in the alar plate (i.e., the sensory part (dorsal half) of the neural tube of the rhombencephalon of primitive jawless vertebrate such as myxinoids (hagfish) or petromyzonts (lampreys; Larsell, 1967; Sugahara et al., 2016). The cerebellum is hence ideally located to accommodate multimodal sensory inputs (Larsell, 1967) including both exteroceptive (lateral-line, vestibular, acoustic, visual) and interoceptive (somatosensory) inputs, thereby functioning as the hub of sensory integration.

In humans, the cerebellum is estimated to share no less than 80% of the CNS neurons in no more than 10% of the brain volume (Herculano-Houzel, 2009). To summarize, the cerebellum has been, throughout its long phylogenetic history, the unique hub in the CNS to accommodate and integrate both afferent and efferent inputs from almost the entire brain, a prerequisite for a neural substrate for adaptive and flexible behaviors.

Nice:

The cerebellum is homogenous in its local neural circuity and heterogeneous in its input-output organization (Ito, 1984). The local neural circuitry of the cerebellar cortex is characterized by its superb homogeneity, sometimes expressed as being ‘‘crystal-like,’’ whereas the inputs to and the outputs from the cerebellum are heterogeneous from one region to another. Therefore, it is commonly postulated that the functional diversity of the cerebellum originates from heterogeneous input-output connectivity and is processed through a common algorithm that is implemented in the homogenous neural circuitry in the cerebellar cortex. In light of Marr’s three levels of analysis, the cerebellum may perform common computation on diversified representations of cerebellar inputs and outputs. Therefore, the cerebellum may be better understood if we ask ‘‘how the cerebellum computes’’ than ‘‘what the cerebellum computes.’’

Among a range of proposals for the cerebellar algorithm including timing processing (Ivry et al., 2002; Ivry and Spencer, 2004) and temporal pattern generator (Fujita, 1982; Dean et al., 2010), one plausible candidate is the prediction of sensory outcomes as a consequence of motor action, referred to as the computation of an internal forward model (Jordan and Rumelhart, 1992; Wolpert and Miall, 1996; Bastian, 2006; Ishikawa et al., 2016).

Visual signals, for example, arrive at the primary visual cortex about 30 ms later and at the parietal cortex about 80 ms later than an onset of a visual stimulus (Schmolesky et al., 1998). Delays in sensory feedback originate from several factors, as quantified in locomotor reflex movements in terrestrial animals of various sizes (More et al., 2010; More and Donelan, 2018). Among factors contributing to the feedback delay such as a synaptic delay or an electro-mechanical delay, the dominant factor is the nerve conduction delay, ranging about 10 ms for a shrew to about 100 ms for an elephant. Larger animals experience a longer feedback delay yet move more slowly, whereas smaller animals experience a shorter feedback delay but move more quickly. In sum, sensory delays are comparable to typical time scales of rapid movements and hence not negligible both in small and large animals.

One mechanism proposed to cope with the delay in sensory feedback is to compute a future state of the body based on a current estimate of the body and an efferent signal of motor control (Figure 1A). This predictive computation internally emulates or models an actual movement of the body by essentially solving an equation of motion of the body forward in time, thereby known as an internal forward model (Wolpert et al., 1998; McNamee and Wolpert, 2019).

They don't quite commit to the cancellation version here..

A clever experiment revealed that the prediction error, but not the target error, drives motor adaptation in a visuomotor rotation experiment by dissociating a prediction error from a target error (Mazzoni and Krakauer, 2006).

The core element in the OFC model is the computation of two gain matrices (Kalman gain and feedback gain matrices). First, the Kalman filter is a recursive method to estimate the current state that integrates a predicted state from an internal forward model and sensory afferents. An optimal estimator is a weighted sum of a predicted state and sensory afferents determined by Kalman gain according to their relative accuracies (or variances). The computation of Kalman gain requires the accuracy of a predicted state from an internal forward model for the optimal tradeoff.

The DN is the final output station from the cerebro-cerebellum. To contribute to the three roles of an internal forward model mentioned above, DNCs should be able to generate dynamic patterns of output. Our previous study unveiled a simple mechanism to explain a wide range of modulation (i.e., facilitation and suppression) of DNC activity (Ishikawa et al., 2014) and solved the controversy over the generation of burst activity of DNCs before limb movement (Thach, 1970; Strick, 1983; Wetts et al., 1985; Chapman et al., 1986; Fortier et al., 1989; van Kan et al., 1993; Goodkin and Thach, 2003) without a major excitatory drive.

To explain the excitation of DNCs [specifically, deep cerebellar nuclei (dCN) cells], two physiological mechanisms have been proposed. One mechanism is the recruitment of a post-inhibitory rebound excitation (Aizenman and Linden, 1999; Hoebeek et al., 2010; Tadayonnejad et al., 2010; Witter et al., 2013). Another mechanism is the suppression of PCs that facilitates dCN cells by a release from tonic inhibition from PCs, a mechanism known as disinhibition (Albus, 1971; Miyashita and Nagao, 1984; Nagao, 1992; Shinoda et al., 1992; Medina and Mauk, 2000).

To address how DNCs become excited or inhibited during voluntary limb movements, we compared the temporal patterns of activity for PCs and DNCs recorded from the same monkeys during step-tracking movements of the wrist (Ishikawa et al., 2014). If the post-inhibitory rebound excitation serves for the facilitation, phasic excitation of PCs and a concomitant inhibition of DNCs should precede the main excitation of DNCs. On the other hand, if the disinhibition serves for the facilitation, we should observe the suppression of PCs and the activation of DNCs at the same timing. We found that (Ishikawa et al., 2014) the majority of PCs in the Cerebro- cerebellum demonstrated suppression before the onset of wrist movements. At the same time, the majority of DNCs were activated without prior suppression.

This finding supports the disinhibition mechanism that the movement-related activation of DNCs occurs when they are released from tonic inhibition from PCs.

In line with our proposal, Dean and Porrill (2010) proposed that the parallel pathways perform in a competitive way to suppress or facilitate the activity of PCs in the cerebellar cortex. Our finding extends their idea to explain the role of the parallel pathways for the generation of dynamic output from DN (Ishikawa et al., 2014).

TanakaIshikawaKakei19

Implements sensory cancellation model, tests with real data:

We here provide neural evidence that the cerebellar circuit can predict future inputs from present outputs, a hallmark of an internal forward model. Recent computational studies hypothesize that the cerebellum performs state prediction known as a forward model. To test the forward-model hypothesis, we analyzed activities of 94 mossy fibers (inputs to the cerebellar cortex), 83 Purkinje cells (output from the cerebellar cortex to dentate nucleus), and 73 dentate nucleus cells (cerebellar output) in the cerebro-cerebellum, all recorded from a monkey performing step-tracking movements of the right wrist. We found that the firing rates of one population could be reconstructed as a weighted linear sum of those of preceding populations. We then went on to investigate if the current outputs of the cerebellum (dentate cells) could predict the future inputs of the cerebellum (mossy fibers). The firing rates of mossy fibers at time t + t1 could be well reconstructed from as a weighted sum of firing rates of dentate cells at time t, thereby proving that the dentate activities contained predictive information about the future inputs. The average goodness- of-fit (R2) decreased moderately from 0.89 to 0.86 when t1 was increased from 20 to 100 ms, hence indicating that the prediction is able to compensate the latency of sensory feedback. The linear equations derived from the firing rates resembled those of a predictor known as Kalman filter composed of prediction and filtering steps. In summary, our analysis of cerebellar activities supports the forward-model hypothesis of the cerebellum.

If we assume that MFs represent a current estimate of state and sensory feedback, PCs represent a prediction of state, and DCs represent a filtered state, then the linear equations can be interpreted as Kalman filter as follows.

Collaterals of MFs send excitatory inputs and PCs send inhibitory inputs to DCs, so wMF→DC and wPC→DC were assumed to be nonnegative and nonpositive, respectively.

For the MF → PC connectivity, the weights exhibited an exponential distri- bution rather than a Gaussian (Fig. 10a), indicating that a relatively few MFs contributed dominantly to the recon- struction of each PC. For the MF → DC connectivity, most of the weights from MFs to DCs were zero reflecting the nonnegative constraint, and nonzero weights were distrib- uted exponentially (Fig. 10b, left panel). Similarly, most of the weights from PCs to DCs were zero reflecting the nonpositive constraint, and nonzero weights were distrib- uted exponentially (Fig. 10b, right panel). Therefore, rela- tively small number of MFs and PCs contributed to the reconstruction of firing rates of each DC.

Yes:

These studies rely on an assumption that kinematic and dynamic representations of PCs relate to forward and inverse models, respectively. This assumption, however, does not hold because an internal forward model should include dy- namical variables such as efference copies of motor control signals. Also, disentangling predicted state signals, sensory feedback signals from the periphery, and motor commands is rather difficult because these signals resemble each other [8]. Therefore, a mere correlative comparison of PC activities with one or other behavioral parameters would not lead to conclusive evidence for either of forward or inverse models.

Our analysis revealed that the current DC activity contained predictive information about the future MF activity for a range of time advance. Therefore, our result supports that the cerebellum is capable of compensating the sensory delays of the order of 100 ms, supporting the forward-model hypothesis.

First, this region of the cerebro-cerebellum receives a putative efference copy as well as a strong somatosensory input [14, 15], and these inputs are presumed to be integrated in the cerebellar cortex. Second, the activities of PCs in this region lag behind those of M1 neurons, while they precede the movement onset [15]. The timing of activity is compatible with the idea that it works as a forward model that predicts an outcome of the motor command. As a result, the output of this region of the cerebro-cerebellum may help M1 to generate a suitable motor command for the next moment depending on the predicted consequence before a feedback signal is avail- able for the current motor command. We note that there are in general two input pathways to the MFs: one from the cerebral cortex through the pons and another from the peripheral sen- sory organs. Our single-unit recording did not allow to iden- tify the origin of MFs and thus to discuss what information the MF activities encoded.

Among these re- current connections, the most relevant to this study is the nucleocortical projections from the cerebellar nuclei to the granular layer as MFs [32]. A recent study reported that excitatory output cells in the interposed nucleus provide efference copy signals via MFs to the cerebellar cortical zones and that an eye-blink conditioning training increased the local density of nucleocortical MF terminals [33]. One may suspect, therefore, that the linear prediction from DCs to MFs reported in this study could be attributed to the nucleocortical recurrent connections. We note, however, that this nucleocortical projection per se does not explain the longer time scale of linear prediction up to 100 ms reported in this study. Also, the nucleocortical pathway comprises only approximately 5% of the total of cerebellar MF inputs [34]. We hence expect that recurrent connec- tions in the nucleocortical projections have minor contri- butions to our results of linear prediction from DCs to MFs.

Anatomical studies using an anterograde tracer (HRP-WGA) in cats reported that MFs of cortical origins via pons consist of the main input to the cerebellar cortex, whereas collaterals of those MFs poorly project directly to the dentate nuclei [52, 53]. Therefore, Bthe bulk of infor- mation of cortical origin reaches the cerebellar nuclei only after processing in the cerebellar cortex^ (p. 22 of Brodal et al. [52]). This is consistent with our assumption that DCs receive a predicted state not from MFs but from PCs.

The perceptron model is essentially a static pattern classifier and not designed to handle time-varying, dy- namic inputs. Subsequently, the perceptron model was extended to the adaptive filter model which generates a dynamic response by summing various temporal basis patterns [58]. The adaptive filter model considers a recur- rent circuit among MFs, Golgi cells, and granule cells which generates resonant temporal patterns with various phase leads and lags. Therefore, the adaptive filter model assumes that the activities of PCs result from the interac- tion among MFs, Golgi cells, and granule cells. There is supportive evidence of the adaptive filter model [59, 60]. On the contrary, our finding of linear transformation from MFs to PCs suggests that the recurrent circuit plays a negligible role in generating temporal bases and rather that MFs already contain rich temporal repertoire that in turn drives the activities of PCs.

58 = Fujita82
59 = DeanPorrill14
60 = DeanPorrillEkerotEtAl10

DeanPorrill14

basically talks about delta rule learning, with STDP variant..

Functional consequences of this bidirectional plasticity have been reviewed by Jo ̈rntell and Hansel (2006), including its ability to explain how parallel-fiber stimulation without conjunctive climbing- fiber stimulation can enormously increase the tactile receptive fields of Purkinje cells in the C3 zone of the cerebellum in vivo (Ekerot and Jorntell, 2003; Jo ̈rntell and Ekerot, 2002, 2011).

One computational feature of the decorrelation learning rule is that synapses for parallel fibers carrying noise or signals unrelated to climbing-fiber signals will eventually be driven to zero, and experimental evidence indicates that a high proportion (up to 98%) of parallel-fiber synapses are indeed si- lent (Isope and Barbour, 2002; Jo ̈rntell and Ekerot, 2002; Wang et al., 2000). Since the only way that silent synapses can take part in new learning is via LTP, it is pos- sible that the properties of in vitro LTP rather than LTD will be relevant to in vivo learning. In fact, a further implication of relying on LTP is that there needs to be plasticity in the inhibitory pathway from granule cells to Purkinje cells via molecular-layer interneurons (stellate and basket cells), otherwise tasks such as eye blink conditioning that require a reduction in Purkinje cell firing could not be learnt (Dean et al., 2010; Ekerot and Jo ̈rntell, 2003; Porrill and Dean, 2008). Such plasticity, predicted by Albus (1971), has been described in vivo (e.g., Jirenhed et al., 2013; Jo ̈rntell and Ekerot, 2003, 2011; Jo ̈rntell et al., 2010),

Assumption seems to be that PC is tonically active?

DeanPorrillEkerotEtAl10

Good review of forward model and data

[rcoreilly]

DEC: Delay Eyelid Conditioning

BoeleKoekkoekDeZeeuw10

Instead, extracerebellar regions like the motor cortex or striatum might initiate the CRs (Aou et al., 1992; Gruart et al., 2000).

TODO: lookup above.

Also, read this paper again and add more info..

JirenhedBengtssonHesslow07

Existing data on Purkinje cell activity during conditioning have been ambiguous and contradictory. We reported previously that a small proportion of Purkinje cells in eyeblink controlling parts of the C3 zone exhibit pause responses to the CS in conditioned animals (Hesslow and Ivarsson, 1994). Others have also reported inhibitory responses, but most observed Purkinje cells have been either excited or nonresponsive (Berthier and Moore, 1986; Ko- tani et al., 2003, 2006; Green and Steinmetz, 2005).

GreenSteinmetz05

Many anterior lobe Purkinje cells showed significant learning-related activity after eyeblink conditioning to one or both of the CSs. More Purkinje cells responded with inhibition than with excitation to CS presentation. In addition, when the firing patterns of all conditioning-related Purkinje cells were pooled, it appeared that the population showed a pattern of excitation followed by inhibition during the CS–US interval. Using cholera toxin-conjugated horseradish peroxidase, Purkinje cells in recording areas were found to project to the interpositus nucleus. These data support previous studies that have suggested a role for the anterior cerebellar cortex in eyeblink conditioning as well as models of cerebellar-mediated CR timing that postulate that Purkinje cell activity inhibits conditioned response (CR) generation during the early portion of a trial by inhibiting the deep cerebellar nuclei and permits CR generation during the later portion of a trial through disinhibition of the cerebellar nuclei.

and record- ings of Purkinje cell activity in this area have revealed inhibitory and excitatory patterns of activity that appear to be related to CS or US presentation and CR execution (e.g., Berthier and Moore 1986; Katz and Steinmetz 1997).

Mauk and colleagues have presented data suggesting that another region of the cerebellar cortex, the cerebellar anterior lobe (lobules I–V), may be involved in CR production and timing (Perrett et al. 1993; Perrett and Mauk 1995; Medina et al. 2000).

two different nuclei involved!

BerthierMoore86

Many PCs responded to the US, the most frequently observed response being a burst of simple spikes. PCs in HVI showed a variety of responses to CSs that were related to conditioned responses (CRs). The most frequently observed response was an increase in simple spikes correlated with CRs. The activity of many of these cells antedated CRs by 20-200 ms. A smaller proportion of cells exhibited inhibition of simple spike activity that antedated CRs. The existence of PCs that alter their firing before CRs suggests that they may be causally involved in this behavior, and in this respect they reinforce reports that lesions of HVI or its connections disrupt nictitating membrane CRs.

Although complex spike activity was not generally related to the US or to CRs, a few PCs responded in relation to CRs with only complex spikes. In demon- strating CR-related activity in cerebellar PCs, this study supports theories of cerebellar learning such as those of Marr and Albus.

Furthermore, multiple-unit recordings from the cerebellar deep nuclei have been reported to covary positively with conditioned NM responding (McCormick and Thompson 1984a, 1984b).

Fifty-three of the 77 cells responded to the US. Forty-nine responded with simple spikes, and 4 responded with complex spikes. Of the former, 39 responded to the US with an increase in firing, 5 with a decrease in firing, and 5 with an increase followed by a period of inhibition. Only 3 cells responded consistently to the tonal CSs independent of the presence or absence of CRs.

39 / 53 increase, 5 / 53 decrease - that is a huge difference!

Thirteen of the 22 pre-CR cells increased simple spike firing, 6 decreased simple spike firing, and 3 increased complex spike firing before CR onset. Fourteen of the 22 cells had complex spikes. Fifteen pre-CR cells were in HVI.

PCs that increase firing before the CR are problematic in this causal role because they imply inhibition of deep nuclear cells. In order to initiate CRs, there would have to be an intervening signal inversion somewhere in the premotor pathway. The proportion of cells that showed CR-related decreases in firing might have been greater at some later stage of training.

Actually, no..

McCormickThompson84a

Neuronal recordings from 323 sites were obtained from throughout the lateral cerebellar cortex and deep nuclei with 91 of the recordings coming from the ansiform lobule (Figs. 1 to 3). Of the 323 recordings, 90 (28%) were found to respond in the CS period. These responses were located within the ansiform lobule, the anterior lobe, the D-I nuclei, and portions of the midline (vermal) cortex. Twenty-nine (32%) of the responses within the CS period correlated significantly with the onset of the eyeblink response (r h 0.67, df = 7). A number (22) of such responses were found within the 91 re- cording sites located in the ansiform cortex above the D-I nuclei (Fig. 2). The significant correlations from this region ranged from 0.76 to as high as 0.98, with a mean value of 0.88. The relatively high values of many of these correlations indicate that these responses were highly related to the performance of the learned eyeblink response (see Figs. 1 and 2). The mean onset latency for these neuronal responses was 29.3 f 16.5 msec before the onset of the NM response and ranged from 5 to 59 msec before the onset of the NM response.

A small number of the responses either significantly increased in amplitude (4 of 22) or decreased in amplitude (4 of 22) during behavioral conditioning.

airpuff were present both before and after the
learning of the behavioral response. A small number signifi-
cantly decreased (6 of 42) and no responses significantly in-
creased in amplitude during behavioral training. The average
onset latency of these responses to the airpuff was 4.6 + 1.9
msec. The possibility that these responses possessed significant
auditory components (the airpuff escaping from the outlet
nozzle generates a broad band noisy stimulus) was tested in 13
of the animals by misdirecting the airpuff to above the animal’s
head for the first block of unpaired trials. Of the 24 recording
sites in these animals, the neurons of 1

clei is related to the somatosensory components of the airpuff and/or performance of the uncon- ditioned response. The fact that the onset latencies of these responses (4.6 msec) is shorter than the earliest eyelid EMG (musculus obicularis oculi) responses to the airpuff (7.5 msec, McCormick et al., 1982c) would rule out feedback from the movement as a cause for the initial portions of these responses. The fact that misdirection of the airpuff did not abolish the neuronal response to the airpuff in the other 9 cases indicates that these responses were auditory in nature. Therefore, with- out such a test, similar responses can only be

[rcoreilly]

Directional Saccade Task data from Lisberger lab

Key results:

• in PCs: CS and SS firing are in opposite directions

• error going left = LTD for right-coding PC: this will decrease inhibition / canceling on the right-coding CN from the PC -- this only makes sense in a sensory cancellation framework!

• LTP happens independent of CS activity -- the "background" / "decay" kind of mechanism.

Overall consistent with:

• IO can only code one direction of error: when sensory signal is stronger than the prediction.
• IO activity can only drive LTD, not LTP.
• Therefore, IO affects the opposite (complementary / competing) coding neurons, not its "own" consistent-coding ones.

This is kind of crazy consistent with everything, but also just plain kind of crazy.

In general, final output is competition between opponent CN outputs -- more efficient to do LTD in opponent pathway in terms of spiking SNR rather than trying to increase an already high spiking rate? This makes a lot of sense.

What about the upgoing and downgoing zebrin stuff? could just be different baserates? why didn't they see these guys in this paper? maybe they did and that was just the other guys?

Also, overall it seems perhaps to be a bit of a stochastic exploratory learning mechanism -- some models have this..

MedinaLisberger08

This is a key methodological and initial results paper, showing correlations of PC spiking (CS and SS) on behavior.

NOTE: data is from one PC neuron! this one neuron shows both ON and OFF modulations (bidirectional).

• a shows that there is a "null" orthogonal direction for each PC where it doesn't respond at all, and if the saccade goes closer to preferred direction, then it fires more (purple) and if away, fires less than null direction (green).

• b shows an OFF-direction case on probe trials (key point): actually does make an OFF direction eye movement, and PC neurons respond accordingly.

• c shows ON case, inverse of above.

• training timing is at 250 ms line shown -- learning is slightly anticipatory.

Group 1 show strong CS modulation during learning: note, only for OFF case!

Group 2 are just weaker version of Group 1 ?

According to IO error logic in sensory terms, this means that IO is getting more sensory inputs during OFF case vs. prediction from CNi -- i.e., it is detecting visual activity in the OFF direction -- this is an "opponent" sensory signal.

This should lead to LTP of CNi to reduce IO CS firing over time, which is observed:

Overall summary data:

Basically, you only get LTD for OFF CS Group 1 guys, and everyone else goes up..

It really seems like a lateral competition story would make a lot more sense.. need to look into that.

YangLisberger14a

Firing rate showed a strong increase for rightward pursuit, a decrease for leftward pursuit, and very slight increases for upward and downward pursuit. We would say that the simple-spike firing of this Purkinje cell had a rightward “on direction” and a leftward “off direc- tion.” Most Purkinje cells in the floccular complex had on directions for simple-spike responses that were toward the side of recording or down- ward (Krauzlis and Lisberger, 1996).

Almost all responsive Purkinje cells had CS responses with direction selectivity opposite to the simple- spike responses (Stone and Lisberger, 1990b; Medina and Lis- berger, 2008). Because we are studying effects linked to CS responses, we will refer to Purkinje cells for the rest of the paper according to the direction tuning of their CS responses: learning instructions will be either in the “ON-CS direction” or “OFF-CS direction.” Note that the ON-CS direction was the off direction for simple-spike responses, and vice versa (Fig. 1B).

This is critical: cells are coding rightward sensory motion, CS error is for left, when leftward sensory motion is greater than predicted -- these are "opponent" cells of the IO / CS signal, per above.

wow crazy short retention: gone in 6s!

As we have noted before (Medina and Lisberger, 2008; Yang and Lisberger, 2013), the difference in the trial-over-trial learned eye velocity between 1– 0 and 0 – 0 pairs implies that the presence and absence of CS responses tends to be correlated across the Purkinje cell population. If CS responses occurred with independent probabilities in each Purkinje cell, then the fluctua- tions across the population in trial-over-trial depression would average out; the trial-over-trial learned eye velocity would be the same for 1–0 and 0–0 pairs.

Maruta et al. (2007) discovered that CS responses often oc- curred more frequently than the average of �1 CS/s. In our data for an instruction–test interval of 2.5 s, 8.5% of learning trials contained �1 CS response during the 400 ms instruction. The magnitude of trial-over-trial depression of simple-spike re- sponses, however, did not depend on whether an instruction caused one versus two CS responses.

NO diff in doublets here!

Figure 4 suggests that long-term learning in pursuit is fa- cilitated by a neural correlate of CS-linked trial-over-trial learn- ing. The directional organization of the simple-spike and CS responses of floccular Purkinje cells (Stone and Lisberger, 1990b;

For Purkinje cells that prefer horizontal pursuit, CS responses will be evoked by rightward or leftward target motion in the left or right floccular complex, respec- tively. Trial-over-trial depression will oc- cur only in one floccular complex, long- term learning would be facilitated in one floccular complex at a time, and each side would work individually to create a learned eye movement of moderate amplitude.

RH = leftward motion, LH = rightward motion. Up and down are in both!

The absence of a large population of
Purkinje cells with CS responses for
downward instructions (Stone and Lis-
berger, 1990b; Krauzlis and Lisberger,

1996. could explain the asymmetry be-
      tween upward and downward trial-over-
      trial learning (Yang and Lisberger, 2010).
      Downward instructions will cause very little CS-linked trial-over- trial simple-spike depression in the floccular complex. The ab- sence of downward trial-over-trial plasticity may deprive the system of the facilitation provided by trial-over-trial learning, and could explain (1) the almost identical downward learning curves for instruction–test intervals of 2.5 and 6 s (Fig. 4C) and (2) the larger upward long-term learning for the short instruc- tion–test interval. We conclude that CS responses may play a primary role in the neural changes that lead to long-term learn- ing, and that the floccular complex may be the site where plastic- ity occurs earliest during pursuit learning.

For Purkinje cells that preferred vertical eye motion (Fig. 5, bot- tom row of graphs), the conclusions are the same. However, the asymmetry in eye velocity learning for upward versus downward instructions (Figs. 4 D, E, 5F ) rendered the data more complicated.

We further suggest that behavioral learning in horizontal pur- suit is symmetrical because of the reciprocal organization of the Purkinje cells in the floccular complexes on the two sides of the cerebellum. We think that the asymmetry of the time courses of neural learning for the ON-CS and OFF-CS directions reflects simple-spike depression that occurs more quickly than does potentiation.
Finally, we suggest that learning in vertical pursuit shows an up–down asymmetry throughout 100 learning trials be- cause of the absence of Purkinje cells with ON-CS directions that are downward. The asymmetry is most pronounced early in a repeated-direction learning block because of the absence of CS-linked trial-over-trial depression for downward instruc- tions. We suggest that the residual asymmetry in simple-spike learning after normalization again reflects quicker simple- spike depression versus potentiation.

In the context of Equation 2, our data show that sensitivity is much higher in Purkinje cells for learned changes in eye velocity than for normal, baseline pursuit (Fig. 6A,C). For instruction directions that cause a high probability of CS responses, learned simple-spike firing rate far exaggerates the learned eye movement on the first learning trial, and remains �2.5 times larger than predicted on the 100th trial.

Thus, we conclude that CS-frequent and CS-infrequent Purkinje cells show different signs of learned changes in simple-spike fir- ing during 100 trials of learning, at least in the ON-CS direction. The learning of the CS-infrequent Purkinje cells in the ON-CS direction adds to the evidence that learned depression of simple-

In a previous paper, we provided evidence that same-trial facili- tation reflects the degree of control of inferior olive firing by the disynaptic pathway from Purkinje cells to the inferior olive (Yang and Lisberger, 2013).

We have argued before for the CS-frequent Purkinje cells that the link from the presence of a CS on the instruction trial to much larger trial-over-trial learning in the behavior is caused by correlations in the CS responses across the Purkinje cell population (Yang and Lisberger, 2013; Yang and Lisberger, 2014). For the CS- infrequent Purkinje cells, the trial-over-trial learning in eye velocity is independent of the occurrence of a CS on the in- struction trial (Fig. 8E). We suggest that the absence of corre- lations in their CS responses accounts for the absence of a link from the CS of the CS-infrequent Purkinje cells to the size of trial-over-trial behavioral learning.

We suggest that CS-infrequent Purkinje cells show “wrong-way” learning in the ON-CS direction because synaptic potentiation on most trials dominates over the small amount of depression that can be mustered on the 20% of trials that show a CS response to the instruction.

(1) long instruction–test intervals do not support trial-over-trial depression and learning and lead to re- duced long-term learning; and (2) the effect of instruction–test interval on long-term learning did not appear for downward in- structive changes in target direction, in agreement with the ab- sence of a large number of Purkinje cells that could undergo CS-linked trial-over-trial depression for downward instructions (Krauzlis and Lis- berger, 1996).

However, trial-over-trial learning and
long-term learning might be controlled in
parallel by the same neural signals, with-
out a causal link from trial-over-trial
learning to long-term learning. In addi-
tion, some long-term learning occurs
without any trial-over-trial effects. Long-
term learning in the cerebellar cortex may
be only partially dependent on CS-linked trial-over-trial depres- sion, or plasticity might occur at other sites in the pursuit circuit.

YES

We conclude that the Purkinje cells with ON-CS directions are carrying 10-times their normal burden in terms of driving eye movement. They may even be the sole site of trial-over-trial plasticity and may be solely responsible for the learned changes in eye velocity.

Indeed, our data suggest that learning involves potentiation in Purkinje cells whose ON-CS directions are opposite to the direction of the instruc- tion. However, potentiation seems to evolve more slowly than depression in vivo and is somewhat weaker.

Thus, the wrong-way learning in CS-infrequent Pur- kinje cells appears to be a consequence of the organization of their climbing fiber inputs, not of their intrinsic physiology.

In the vestibulo-ocular reflex, plasticity occurs in both the cerebellar cortex and the deep cerebellar nucleus (Lisberger, 1994; Blazquez et al., 2006). There are two components of learn- ing that do versus do not depend on climbing-fiber inputs (Ke et al., 2009). Learning occurs on multiple time courses (van Alphen and De Zeeuw, 2002). Vestibulo-ocular reflex learning to in- crease versus decrease the size of the reflex employs different cellular mechanisms of plasticity (Boyden et al., 2006).

YangLisberger14b

Plasticity and cerebellar motor learning are graded by the duration of complex-spikes

Yep clearly an effect!

YangLisberger17

In this article, we take advantage of our ability to quan- tify single-trial neural and behavioral learning to show four new features of cerebellar learning.

First, modulation of the context of instructive target motions shows new cor- relative parallels between the probability and duration of CS responses and the resulting neural and behavioral learning.

Second, when the system is in a stationary learning condition, neither the probability nor duration of CS responses fluctuates slowly over the course of a learning session.

Third, the different time courses of syn- aptic depression and potentiation in the presence versus absence of a CS response suggest the existence of inde- pendent mechanisms.

This suggests that it is not a competitive interaction.

Fourth, comparison of the predic- tions of computer simulations to analyses of real data constrains a biologically motivated model of trial-over-trial learning that can account for the statistics of a complex set of data.

In our prior article (Yang and Lisberger, 2014a), we argued that the modulation of CS duration occurred in the inferior olive because of a lack of correlation between CS duration and simple-spike firing rate, the latter an indicator of the state of the Purkinje cell membrane. If our analysis had been able to constrain the degree of underlying correlation of CS probability and dura- tion, we might have been able to shed some light on the site of the modulations.

[rcoreilly]

VOR paradigm

KimpoRinaldiKimEtAl14

Despite the considerable evidence that climbing fiber activity provides error signals controlling
motor learning, several studies have called this idea into question. Recordings during some cerebel-
lum-dependent tasks have revealed a dissociation between the encoding of errors and the induction
of learning over the course of a training session (Catz et al., 2005; Ke et al., 2009; Kitazawa et al.,
1998; Ojakangas and Ebner, 1992). In addition, perturbation of the most extensively studied form of
climbing fiber-triggered plasticity, long-term depression of the parallel fiber-to-Purkinje cell synapses
(pf-Pk LTD), impairs certain learning tasks while leaving other cerebellum-dependent learning tasks
intact (Aiba et al., 1994; Boyden et al., 2006; De Zeeuw et al., 1998; Feil et al., 2003; Hansel et al.,
2006; Katoh et al., 2000; Katoh et al, 2005; Kim and Thompson, 1997; Kishimoto et al., 2001;
Kishimoto et al., 2001; Koekkoek et al., 2003; Schonewille et al., 2011; Shibuki et al., 1996;
Shutoh et al., 2002, 2003; van Alphen and De Zeeuw, 2002; Welsh et al., 2005; Yanagihara and
Kondo, 1996). Thus, after decades of research, the conflicting evidence has left unresolved the ques-
tion of whether error signals in the climbing fibers are the driver of motor learning.

However, more recently, it has been shown that a
Purkinje cell’s response to its climbing fiber input can be graded, which, in turn, can regulate the effi-
cacy of climbing fiber stimulation to induce plasticity in vitro and in decerebrate preparations (Weber
et al., 2003; Carey and Regehr, 2009; Maruta et al, 2007; Mathy et al., 2009; Rasmussen et al.,
2013).

Previous efforts to determine whether these error signals in the climbing fibers are what drive
VOR learning have been inconclusive. Over the course of both VOR-increase and VOR-decrease
learning, the Purkinje cell simple spike output during the VOR is altered in a manner consistent with
the induction of LTD in parallel fiber inputs that were coactive with the climbing fibers during training

(Dufossé et al., 1978; Hirata and Highstein, 2001; Lisberger et al., 1994; Miles et al., 1980; Nagao,
1989; Watanabe, 1984, 1985). Also, direct, optogenetic activation of the climbing fibers can induce
VOR-increase learning, when paired with a vestibular stimulus (Nguyen-Vu et al., 2013). On the other
hand, a previous attempt to induce VOR-decrease learning with climbing fiber stimulation was unsuc-
cessful (Nguyen-Vu et al., 2013). Moreover, studies of multiple lines of mice deficient in pf-Pk LTD
have not consistently found impairments of VOR learning, and the reported impairments are more
pronounced for VOR-increase than VOR-decrease learning (Aiba et al., 1994; Boyden et al., 2006;
De Zeeuw et al., 1998; Feil et al., 2003; Hansel et al., 2006; Kim and Thompson, 1997; Koekkoek
et al., 2003; Schonewille et al., 2011; van Alphen and De Zeeuw, 2002). These results raised the
possibility that the efficacy of the error signals carried by climbing fibers to induce plasticity is reduced
during VOR-decrease training. In this study, we provide convergent evidence for this possibility from
recording and stimulation experiments. The regulation of climbing fiber efficacy for inducing plasticity
represents a new component of the learning algorithm implemented by the cerebellum.

KEY point: VOR decrease much less trainable than increase. Unidirectional learning in PC, again!

The climbing fiber input to Purkinje cells in the flocculus encodes the retinal slip that drives VOR
learning (Simpson and Alley, 1974; Ghelarducci et al., 1975; Graf et al., 1988; Stone and Lisberger,
1990). Retinal slip can induce an adaptive increase or decrease in the gain of the VOR, depending on
its direction relative to the vestibular stimulus (Boyden et al., 2004). To induce a learned increase in
VOR gain, visual image motion is paired with a vestibular stimulus in the opposite direction (Figure 2A);
to induce a learned decrease in VOR gain, visual image motion is paired with a vestibular stimulus in
the same direction (Figure 2B; Collewijn and Grootendorst, 1979; Vercher and Gauthier, 1990–1991;
Pastor et al., 1994; Raymond and Lisberger, 1996). Most climbing fibers in the flocculus increase
their firing in response to contraversive retinal slip (image motion away from the side on which the cell
is recorded) (Simpson and Alley, 1974; Raymond and Lisberger, 1998; Stone and Lisberger, 1990;
Figure 2C–F). During VOR-increase training, the contraversive retinal slip that activates climbing fibers
occurs during ipsiversive vestibular stimuli (Figure 2A,C,E and Figure 2—figure supplement 1A,C,E),
whereas during VOR-decrease training, the contraversive retinal slip and associated increase in
climbing fiber activity occur during contraversive vestibular stimuli (Figure 2B,D,F and Figure 2—
figure supplement 1B,D,F). The same climbing fibers encode retinal slip during both training para-
digms, and the amplitude of the response is the same (Figure 2E,F). What distinguishes the two
paradigms, and carries information about the required direction of learning, is whether the increased
climbing fiber activity occurs during vestibular stimuli in one direction vs the other.

Such changes in Purkinje cell responses have been reported by several laboratories after VOR learning (Dufossé et al., 1978; Hirata and Highstein, 2001; Lisberger et al., 1994; Miles et al., 1980; Nagao, 1989; Watanabe, 1984, 1985). However,
the question of whether the error signals in the climbing fibers are what triggers these changes has
been controversial (Ke et al., 2009; Nguyen-Vu et al., 2013; reviewed in Ito, 1982; Lisberger, 1988).

Climbing fibers fire at a low rate, and even when retinal slip is present, an individual climbing fiber does
not fire on every trial. We harnessed this natural variance to assess whether the activation of the
climbing fibers during training is what drives plasticity in the VOR circuit. If a spike in the climbing fiber
on one trial induces plasticity in the inputs to its Purkinje cell target, this might be detected as a
change in the Purkinje cell’s simple spike response on the next trial, as previously reported during
smooth pursuit eye movement learning (Medina and Lisberger, 2008; Yang and Lisberger, 2010,
2013).

During VOR-increase training, there was a tight, trial-by-trial
correlation between the activity of the climbing fiber input to a given Purkinje cell and the induction
of changes in its simple-spike output (Figure 3B). If there was a spike in the climbing fiber input on the
first trial of a pair, there was a reduction of about 8 spikes/s in the firing rate of the Purkinje cell on the
subsequent trial (Figure 3B,C; CF–No CF, red). On trials with no climbing fiber spike, there was no
reduction in Purkinje cell firing rate on the subsequent trial (Figure 3B,C; No CF–No CF, black). The
sensory error (retinal slip speed) was indistinguishable on trials with a climbing fiber spike vs without a
climbing fiber spike, and therefore cannot account for the difference observed on the trial after the
climbing fiber spike (Figure 3—figure supplement 1).

The Purkinje cells did undergo gradual, adaptive changes in their responses over the course of the
full VOR-decrease training session, which could be detected during the same ∼90-s VOR-decrease
training sessions used for the trial-by-trial analysis (Figure 4). Across trials of VOR-decrease training,
there was a progressively lower Purkinje cell firing rate during contraversive vestibular stimuli, similar
to what has been reported previously using much longer training periods (Dufossé et al., 1978;
Hirata and Highstein, 2001; Lisberger et al., 1994; Miles, Braitman, et al., 1980; Nagao, 1989;
Watanabe, 1984, 1985). This observation, along with the observation that lesions of the floccular
complex disrupt both VOR-increase and VOR-decrease learning (Ito et al., 1982; Koekkoek et al.,
1997; Lisberger et al., 1984; Nagao, 1983; Rambold and Churchland, 2002), indicates that the
Purkinje cells in our sample participate in VOR-decrease learning as well as VOR-increase learning.
However, there was no trial-by-trial correlation between the climbing fiber activity and the plasticity of
the Purkinje cell responses to suggest a causal relationship during VOR-decrease training, as there was
during VOR-increase learning. Thus, although the performance errors elicited equally robust responses
in the climbing fibers during the two learning paradigms, those responses seem to have a different
impact in terms of their ability to induce plasticity.

• PC plasticity happens
• It doesn't depend on IO CS activity
• IO CS is the same for both directions -- that part is a bit weird.
• Also, shouldn't we be looking in the opposite hemisphere or something?

When the number of spikelets, and hence the complex spike
duration, is manipulated in vitro or in decerebrate animals (Carey and Regehr, 2009; Mathy et al.,
2009; Rasmussen et al., 2013), the probability of climbing fiber-induced plasticity can be altered, with
shorter complex spike durations associated with a lower probability of climbing fiber-induced plas-
ticity. We evaluated whether shorter complex spike durations could explain the lack of a trial-to-trial
effect of climbing fiber activity during VOR-decrease training, by analyzing the waveform of complex
spikes during each training paradigm (Figure 2E,F), in the same analysis windows used for the trial-by-
trial analysis. Surprisingly, we found that complex spike waveforms during VOR-decrease training had
more spikelets and were of longer duration than those during VOR-increase training (Figure 5A).

Thus, the reduced efficacy of climbing fibers to induce plasticity during VOR-decrease learning cannot be
explained by a shorter complex spike duration.

It is conceivable that climbing fiber-triggered plasticity would be less effective at reducing Purkinje
cell simple spike firing if the firing rate was already very low. Therefore, we compared the average
simple spike responses during vestibular stimuli in the direction associated with elevated climbing
fiber activity: ipsiversive and contraversive vestibular stimuli during VOR-increase and VOR-decrease
training, respectively. For both, firing rate decreased relative to baseline during the training stimulus
(Figure 2G,H), but this reduction was smaller during VOR-decrease training than VOR-increase training
(Figure 5C). This suggests that there is no floor effect limiting plasticity of the simple spike responses
during VOR-decrease training.

Nevertheless, the difference in simple spike rates during the two
training paradigms suggests different patterns of excitatory and inhibitory input to Purkinje cells, which
might regulate climbing fiber-triggered plasticity.

We recently reported preliminary evidence that climbing fiber stimulation can
induce VOR-increase but not VOR-decrease learning (Nguyen-Vu et al., 2013). However, the single
set of stimulation parameters used in that study was designed to elicit a maximal response, and
could potentially have recruited additional plasticity that masked the expression of any climbing
fiber contribution to VOR-decrease learning. Moreover, there was no test of whether climbing fiber
activation could be initiating changes in the VOR circuit that would support the delayed expres-
sion of VOR-decrease learning at later time points beyond the training session, which is plausible
given the evidence that different mechanisms can support oculomotor learning over different time
scales (Titley et al., 2007; Shutoh et al., 2006; Okamoto et al., 2011). Here, we address those
limitations by testing a broader range of stimulation parameters and by measuring learning 2 hr
after training as well as during the 30-min training session (Figures 6 and 7).

If the climbing fibers were stimulated during the ipsiversive phase of the same
vestibular stimulus, the VOR gain was higher than the vestibular-only control at the end of training
(Figure 6B, red and Figure 7A, pink). This effect of climbing fiber stimulation was long-lasting; when
retested 2 hr after training, the VOR gain was still higher compared to vestibular-only training (Figure 6B,
red and Figure 7B, pink). For both training paradigms, there was an increase in VOR gain between the
end of training and the 2 hr retest, presumably reflecting decay of the non-associative habituation
(Figure 6B, black and Figure 7, dark grey).

To roughly mimic the pattern of climbing fiber activation during VOR-decrease training (Figure 2B,D,F)
(Ke et al., 2009; Raymond and Lisberger, 1998; Simpson and Alley, 1974; Watanabe, 1984, 1985),
the climbing fibers were optogenetically activated during the contraversive phase of the vestibular stim-
ulus (Figure 6A, blue). This pairing had no detectable effect, immediately or 2 hr post-training. There
was no significant reduction of the gain of the VOR below the habituation level observed in response
to training with the vestibular stimulus alone (Figure 6B, blue and Figure 7, light blue). In contrast, when
the vestibular stimulus was paired with an appropriate visual stimulus (see ‘Materials and methods’),
there was a bigger decrease in VOR gain than induced by the vestibular stimulus alone, demonstrating
that an associative decrease in VOR gain was possible (Figure 7A, white). Therefore, we tested whether
stronger climbing fiber stimulation could induce an associative decrease in VOR gain.

Key point: lack of decrease is behavioral here.

We stimulated the climbing fibers three times during each cycle of the vestibular stimulus (125 ms
inter-stimulus interval), either unilaterally or bilaterally. These stronger climbing fiber stimulation para-
digms induced bigger increases in VOR gain when delivered during the ipsiversive phase of the ves-
tibular stimulus (Figure 7, dark pink and red). However, when timed to induce VOR-decrease learning,
increasing the number of climbing fiber stimuli had no effect (Figure 7, blue and dark blue).

BoydenKatohRaymond04

Although the function of the VOR is to support clear vision, the VOR is measured in the dark to isolate the eye movements driven by vestibular stimuli from eye movements driven by visual stimuli. The performance of the VOR is characterized by the gain, which is defined as the ratio between eye and head velocities. If the gain of the VOR is poorly calibrated, then head movements result in image motion on the retina, resulting in blurred vision. Under such conditions, motor learning adjusts the gain of the VOR to produce more accurate eye motion. Such adjustments are needed throughout life, as neurons and muscles develop, weaken, and die—or for humans, when a new pair of eyeglasses changes the magnification of the visual field. In the laboratory, we induce motor learning in the VOR by pairing image motion with head motion. Depending on the relative direction of head motion and image motion, the gain of the VOR can be adaptively increased or decreased. An increase in VOR gain is induced by image motion in the direction opposite that of the head (gain-up stimulus), and a decrease in VOR gain is induced by image motion in the same direction as the head (gain-down stimulus) (Figure 2).

Key point: all these figures just have one signal going into the IO -- completely not getting point of IO as comparator of the very same vestibular AND ocular signals being regulated!!

Ito proposed an implementation of the Marr-Albus hypothesis (Ito 1972, 1982), in which the role of the cerebellum is to store the motor memory for the learned change in VOR gain.

An alternative model was proposed by Miles & Lisberger (1981). They proposed that the role of the cerebellum was not to store the motor memory but rather to compute the instructive signal guiding the induction of plasticity. According to this model, Purkinje cells convey the instructive signal to the vestibular nucleus, where it triggers heterosynaptic changes in the connections between vestibular afferents and neurons in the vestibular nucleus. The Miles- Lisberger model attributes the altered responses of Purkinje cells to altered input to the cerebellum from mossy fibers carrying an efference copy of the altered eye movement command created in the vestibular nucleus. Thus the Purkinje cells could exhibit learning-related changes in their responses during performance of the VOR, even if no synaptic changes occur within the cerebellum. In contrast, the Marr-Albus-Ito model attributes altered Purkinje cell responses, recorded during performance of the VOR, to synaptic changes onto Purkinje cells themselves.

Propose changes in MF inputs, due to adaptation in VN. This is more consistent with nuclear model.

Thus two long-standing hypotheses propose (a) different sites of plasticity for motor learning, (b) different instructive signals guiding plasticity at these sites, and (c) different explanations for the changes recorded in Purkinje cells during performance of the VOR. Experimental tests of these three predictions provide support for and against each of these two hypotheses.

Measured in both of these ways, the sensitivity of Purkinje cells to vestibular input changes in the direction opposite to that predicted by the action of cerebellar LTD, and in the direction opposite that required to account for the change in VOR gain, given the known connectivity of Purkinje cells with their downstream targets (Hirata & Highstein 2001, Lisberger et al. 1994a, Miles et al. 1980). This result is difficult to reconcile with the idea that plasticity of the vestibular inputs to Purkinje cells is the primary mediator of the learned change in the gain of the VOR. Although some investigators have raised concerns about the assumptions associated with com- paring neural responses across these different behavioral conditions (Ito 1993, Tabata et al. 2002), further comparisons between neural signals recorded in a sin- gle behavioral condition circumvent these concerns and provide sound evidence for extracerebellar plasticity.

Yay CN!

NO is involved in a form of parallel fiber–Purkinje cell long-term potentiation (LTP) (Lev-Ram et al. 2002), and it regulates the intrinsic spiking of Purkinje cells (Smith & Otis 2003), modulates synapses onto granule cells (Wall 2003), and alters processing in the vestibular nuclei (Moreno-Lopez et al. 2002).

In the Miles-Lisberger model the neural instructive signal is carried by Purkinje cells, which control the induction of plasticity in the vestibular nucleus.

For a neuron to provide an instructive signal for the VOR, its patterns of spiking must discriminate stimuli that increase and decrease VOR gain. During training with stimuli typically used to induce learning (low frequencies of head and visual stimulus oscillation), both climbing fiber and Purkinje cell responses discriminate the gain-up and gain-down stimuli (Simpson & Alley 1974, Watanabe 1984). Therefore, either population of neurons could, in theory, trigger the differ- ent changes required under these two conditions, and hence these findings are compatible with both the Marr-Albus-Ito and Miles-Lisberger models. However, during the induction of learning with high-frequency training stimuli, only climb- ing fibers, not Purkinje cells, discriminate the stimuli that increase or decrease VOR gain (Raymond & Lisberger 1998). This finding indicates that Purkinje cells cannot trigger the different changes induced by the gain-up and gain-down stimuli at high frequencies. Therefore, at least at high frequencies, the climbing fiber is the best candidate for the instructive signal guiding motor learning in the VOR.

Inputs to the vestibular nuclei undergo both potentiation and de- pression (Caria et al. 1996, 2001; Racine et al. 1986), but whether Purkinje cell spiking, the key instructive signal in the Miles-Lisberger model, can control the induction of plasticity at these synapses is not clear (Babalian & Vidal 2000).

Studies that reported small effects of lesions on the expression of previously acquired changes in VOR gain generally used longer training paradigms than did those that reported large effects.

An early study found that increases in VOR gain passively decayed more rapidly than did de- creases in VOR gain (Miles & Eighmy 1980). Furthermore, increases in VOR gain can be actively reversed more readily than can decreases in VOR gain (Boyden & Raymond 2003). Finally, increases in gain generalize less than decreases when measured in a context different from that used during training (see below). The dis- tinct properties of increases and decreases in VOR gain suggest they are mediated by different plasticity mechanisms.

These are global behavioral differences: not about laterality or different zebrin zones or anything -- just plain different!

Other movements may also depend on different plasticity mechanisms to implement bidirectional changes in movement amplitude. For example, increases and decreases in the gain of saccadic eye movements exhibit different behavioral properties (Robinson et al. 2003, Straube et al. 1997).

In vitro, parallel fiber synapses exhibit not only LTD but also two different forms of LTP. Both forms of LTP are induced when parallel fibers are active in the absence of climbing fiber activity (Lev-Ram et al. 2002, Sakurai 1987, Salin et al. 1996). Thus the learning rule for induction of LTP is complementary to that for LTD: In the terminology of Boolean logic, LTD is induced when parallel fiber activity AND climbing fiber activity occur simultaneously, whereas LTP is induced when parallel fiber activity AND NOT climbing fiber activity, occur.

ALSO, Kimpo et al asymmetries are all on the same neurons, so different upbound / downbound issues do not apply.

In the VOR, an asymmetry exists in the reversal of increases and decreases in VOR gain: Increases in gain are readily reversed by subsequent gain-down training, but decreases in gain are harder to reverse by gain-up training (Boyden & Raymond 2003). This finding suggests that increases and decreases in VOR gain depend on different plasticity mechanisms that reverse each other with unequal efficacy.

A study of the strength of parallel fiber–Purkinje cell synapses in rat cerebellar slices estimates that only 7% of the connections are actually capable of driving postsynaptic responses (Isope & Barbour 2002). Thus in the basal state, there may be little additional room for LTD to navigate, and the greatest capacity for circuit modification could lie in potentiating the large number of silent parallel fiber synapses.

In support of this view, a recent study in vivo found that stimulation of parallel fibers in zone C3 of the cerebellum resulted in expansion of Purkinje cell cutaneous receptive fields, which the authors attributed to the action of LTP on parallel fiber–Purkinje cell synapses (Jorntell & Ekerot 2002). Contraction of Purkinje cell receptive fields became possible only after prior expansion had occurred, which suggests that one function of LTD is to reverse previously induced LTP.

Increases in VOR gain generalize less than decreases in VOR gain across a num- ber of sensory dimensions.

Consistent with specificity of PC / cortex learning. Key point: cortex is all about learning special case overrides / exceptions to the standard rules!

Also, something more happening with inc vs. dec than the IO CS asymmetry, which wasn't even there (?). Kimpo shows that even with optogenetic firing of CSs, doesn't drive LTD for VOR decreases..

VogesWuPostEtAl17

This work from De Zeeuw lab in mice is strangely disconnected from the Raymond lab work above, in monkey. Also does not seem internally consistent.. weird.

Because spiking activity of the climbing fibres is highly direction-selective (Simpson & Alley, 1974), the emerging hypothesis is that plasticity mechanisms contributing to learning are direction-selective too. Hence, cerebellar learning, in terms of mechanism and strength, may in principle also depend on the direction of sensory input and/or desired motor output. However, for adaptation of the vestibulo-ocular reflex (VOR), which is one of the most studied cerebellar motor learning tasks (Lisberger & Fuchs, 1978; Hirata & Highstein, 2000; Ito, 2006; Ushio et al. 2011), the efficacy of direction-specific learning has not been assessed in detail yet and the quantitative relation between movement direction and adaptation still remains to be studied in the same experiments (Ushio et al. 2011; Migliaccio & Schubert, 2013; Nguyen-Vu et al. 2013).

Hello? Kimpo et al? I must be missing something.

More specifically, the contribution of both SS and CS activity of PCs to adaptation of the VOR remains enigmatic (Wulff et al. 2009; Nguyen-Vu et al. 2013; Kimpo et al. 2014; Badura et al. 2016; Titley and Hansel, 2016).

Kimpo seems very non-enigmatic to me.

several predictions can be made on potential direction-dependent plasticity mechanisms, even with data that have been collected following sinusoidal stimulations. Deficits in potentiation mechanisms in PCs robustly impair VOR gain–increase, although with a lesser effect on gain–decrease paradigms (Schonewille et al. 2010), whereas selective deficits in LTD at the parallel fibre to PC synapse do not appear to lead to obvious changes in any form of VOR adaptation (Schonewille et al. 2011)

This is directly contradictory to Kimpo, and to their own results!? gain increase is driven by LTD, gain decrease by LTP.

From Kimpo:

Moreover, studies of multiple lines of mice deficient in pf-Pk LTD
have not consistently found impairments of VOR learning, and the reported impairments are more
pronounced for VOR-increase than VOR-decrease learning (Aiba et al., 1994; Boyden et al., 2006;
De Zeeuw et al., 1998; Feil et al., 2003; Hansel et al., 2006; Kim and Thompson, 1997; Koekkoek
et al., 2003; Schonewille et al., 2011; van Alphen and De Zeeuw, 2002).

Interestingly, in the related Purkinje cells (vertical-axis type; see Methods) increasing SS activity and thus decreasing CS activity is known to correlate with naso-temporal (n–t) eye movements, during both the VOR and optokinetic reflex (OKR) evoked by visual input (Graf et al. 1988; De Zeeuw et al. 1995; Schonewille et al. 2006).

SchonewilleGaoBoeleEtAl11

actually report no effect on 3 different tasks!

For example, blockade of LTD induction by interference with the mGluR1/PKC, PKG, or aCamKII path- ways all resulted in impairment of VOR adaptation (Aiba et al., 1994; De Zeeuw et al., 1998; Feil et al., 2003; Hansel et al., 2006). Still, these studies were not conclusive, because pharma- cological blocking of LTD did not affect eyeblink conditioning (Welsh et al., 2005), and training without instructive signals from the climbing fibers partially allowed VOR adaptation (Ke et al., 2009). In principle the positive correlations found in the mouse mutants in which induction of LTD was affected could be attributed to the fact that the affected receptors and kinases mediate upstream signaling in a highly divergent fashion. Each kinase has many substrates, most of which are not involved in PF-PC LTD and could affect both baseline function of the cerebellar network and other forms of synaptic and nonsynaptic plasticity in the cerebellum (Chen and Tonegawa, 1997; Hansel et al., 2006; Kano et al., 1996).

RaymondMedina18

Following a single climbing fiber response to an error on one trial, there is an approximately 5% change in the output of the Purkinje cell targets of this climbing fiber and in the behavioral response on the next trial (Medina & Lisberger 2008). These changes are roughly 100-fold larger than expected from the cumulative changes that accrue across multiple trials, but they decay quickly, in approximately 6 s (Khilkevich et al. 2016, Kimpo et al. 2014, Yang & Lisberger 2014b). One candidate synaptic mechanism underlying these rapid, transient changes in the neural and behavioral responses is short-term associative plasticity of the granule cell–Purkinje cell synapses (Brenowitz & Regehr 2005, Suvrathan et al. 2016).

most likely because the recruitment of plasticity at the granule cell–Purkinje cell synapses is highly sensitive to the parameters of training (Boyden et al. 2006, Nguyen-Vu et al. 2013). Notably, different Purkinje cells exhibit a correlate of learning at different times within a training session (Yang & Lisberger 2014b), with some individual cells exhibiting changes only transiently for a few dozen trials and then apparently handing off the memory trace to other Purkinje cells (Li et al. 2011). Thus, there appear to be multiple learning time constants and mechanisms of plasticity within the minutes–hours range.

[rcoreilly]

Up / down zebrin parcellation

Hypothesis: zebrin up / down modulation is more associated with skeletal muscle systems that need strong opponent processing -- not so much in vestibular systems.

HawkesLeclerc87

early staining paper

vestibular:

there is a rostrocaudal segregation
of the efferent terminal fields of median vermis Purkinje
cells where they synapse in the fastigial nuclei. The
mabQ113+ cells, probably corresponding primarily to P I +,
tendtoterminateinthecaudalpoleoftheipsilateralfasti- spect,lobuleVIIIanditslateralextensionasthecopulapyr- gial nucleus, whereas the Purkinje cells of P1- terminate amidis is noteworthy in that it is here that all seven P+
in the rostra1 pole (Hawkes and Leclerc, '86).

SugiharaShinoda07

In short, groups I–V may be primarily concerned with motor control that uses cerebral and mesodiencephalic (group I), vestibular (group IIa) and collicular (group IIb), somatosensory plus mesodiencephalic and vestibular (group III), somatosensory (group IV), and visual (group V) information, respectively. Our results suggest that we can extend our group hypothesis to the CN. That is, each compartment identified in the CN is proposed to be involved in the same function that has been proposed for its corresponding olivocerebellar group. To give an example, differ- ent functions such as control of locomotion (Mori et al., 1999), saccade and other eye movements (Ohtsuka and Noda, 1991), and posture and head orientation (Wilden et al., 2002) have been reported in the FN. The subdivisions of the FN into the rostro- dorsal aldolase C-negative part (group III-spinal), centrodorso- caudal part (group IIb-collicular), and centrocaudal part (group IIa-vestibular) in the present study may be directly related to the above functions (locomotion, eye movements, and head orienta- tion). As to the nucleofugal projections from the FN, it has been reported that the caudal FN, which essentially belongs to group II (vestibular and collicular), projects to the vestibular nucleus, but the rostral FN (group III in the present study) does not (Lee et al., 1989; Kurimoto et al., 1995; Teune et al., 2000). Similarly, group- ing all of the CN compartments according to this scheme may be a way to split the CN into functionally distinct motor regions. Therefore, the nuclear compartmentalization defined in the present study may reflect a closer relationship to specific func- tional divisions of the CN with regard to motor control than do the traditional nuclear subdivisions (fastigial, interposed, and dentate nuclei).

SugiharaShinoda04

DeZeeuw21

Main paper

Adaptation of the VOR, which serves to main- tain stability of images on the retina during head movements, is controlled by upbound microzones in the floccular complex of the vestibulocerebellum 40,68,110–113.

upbound only??

During adaptation, when the gain of the ipsiversive VOR has to be further increased with respect to that of head rotation, the simple spike activity of the floccu- lar Purkinje cells increases in an approximately linear fashion over time, at least during the initial stages of learning73,110,117,118 (Fig. 5a). This increase in Purkinje cell activity appears to be sufficient to induce the required behavioural increment, as optogenetic stimulation of the floccular Purkinje cells can drive the eye movements and the amplitude of these eye movements covaries with the intensity of optogenetic stimulation73.

This seems to be an alternate universe to above VOR story?

[rcoreilly]

Functional organization of cerebellum

In general, it is just a big outer product system with lots of stuff going multiple places and getting intermixed with other stuff..

Cerebellar Nuclei (CN)

MantoOuladBenTaib10

The vestibular and FN are concerned with the control of eye movements, control of head orientation, stance, and gait. FN can be func- tionally divided into rostral and caudal components [4, 5]. The rostral portion is involved in the control of somatic musculature, head orientation, and eye-gaze shifts [4].

The caudal FN plays key roles in saccade generation and smooth pursuit [6].

The IN is particularly active during the modulation of various reflexes and sensory feedback [7]. The eyeblink responses are typically associated with a modulation of activity in behaving animals [8]. The intermediate cortex and the IN fire in relation to the antagonist muscle group [9, 10], in agreement with a role in damping the limb oscillations and compensation of errors [11]. The IN participates in the excitability of the stretch reflexes [12].

The DN is especially concerned with voluntary movements of the extremities, including reaching and grasping. Dentate neurons preferentially fire at the onset of movement triggered by mental associations [4].

KebschullCasoniConsalezEtAl24

Although local interneurons as well as local (recurrent) collaterals of projection neurons have been described [69], internuclear connections have not yet been reported as a prominent feature of internal CN organization. An overview of the extracerebellar sources of cerebellar afferents is provided in Table 2 and illustrated schematically in Fig. 5.

As such, it has been speculated that spinal projection patterns to the nuclei resembling the adult organization predate adult terminal patterns in the cerebel- lar cortex. It therefore seems quite possible that the final fine-tuning of cortical CF organization may be based on functional connections made by their parent fibers in the CN.

total number of nuclear neurons in the mouse cerebellum has been estimated at approximately 20,000 [30].

Despite a wealth of literature on the subject, a comprehen- sive description of all the ins and outs of the CN cannot yet be given. This is due to the diversity of CN cell types (see “Cell Types of the Adult Cerebellar Nuclei”) in combination with the complex organization of both the terminal distribu- tion of diverse groups of afferents (Table 2) as well as of the wide and complex distribution of CN efferents (see “Effer- ent Connections of the Cerebellar Nuclei”).

In general, it can be said that systems that deal with rather direct cutaneous or proprioceptive information (e.g., the column of Clarke, the dorsal column nuclei) provide no, or only scant, projections to the CN [46, 49, 88]. Conversely, MF systems originating from other parts of the spinal cord and the medulla (e.g., reticular and vestibular nuclei) are prominent sources of CN innervation (Table 2) [53, 57]. For the sake of simplicity and following past convention, though, we will here refer to the afferents as “MF.”

88 = PopEspinosaBlevinsEtAl22

Analogous to the electrophysiological signatures of neurons in the vestibular nuclei to which the CN are often considered closely related [187–189], the CN neurons were classified into two electrophysiological groups: fast-spiking, large neurons that were assumed to be the principal (projection) neurons of the CN, and smaller, slow-spiking neurons, thought to represent interneurons.

The resulting, at the time somewhat surprising revelation of GABAergic neurons expressing slower spike frequencies than putative glutamatergic neurons, was subsequently complemented by using increasingly specific genetic tools, such as viral transfection in combination with cre-lox expression systems, to differentiate axonal target regions [173, 175].

PopEspinosaBlevinsEtAl22

Interesting: proprioception is a cortical MF input, but NOT a CN input -- not something that is predicted but rather something that is used for prediction.

When proprioception is lost, gross tra- jectories are maintained, but coordinated limb movement is impaired (Gordon et al., 1995; Abelew et al., 2000; Windhorst, 2007; Akay et al., 2014). Muscle and tendon information detected by proprioceptive sensory neurons is integrated by secondary neurons in the spinal cord and relayed to the cerebellum through both direct and indirect spinocerebellar pathways (Oscarsson, 1965; Bosco and Poppele, 2001; Jiang et al., 2015). In this study, we sought to define the precise anatomy of the proprioceptive system through the spinal cord using genetic tools in mice.

A major contributor to the DSCT comes from Clarke’s column (CC) neurons, whose soma reside in the medial aspect of the thoracic to upper lumbar spinal cord (Oscarsson, 1965; Baek et al., 2019). While SCT axons terminate as mossy fiber (MF) terminals on granule cells (GCs) in the cere- bellum (Arsenio Nunes and Sotelo, 1985; Reeber et al., 2011), the extent to which CC neurons send axon collaterals to areas within the spinal cord, medulla, or cerebellar nuclei (CN) is unclear (Ekerot and Oscarsson, 1976; Matsushita and Gao, 1997; Mogensen et al., 2017; Luo et al., 2018). Such axon collaterals would be important for integration with other ascending or de- scending pathways (Sillitoe et al., 2012; Beitzel et al., 2017).

Compared with the direct spinocerebellar pathways, less is known about the anatomy of the indirect spino-lateral reticular nucleus (spino-LRt) and spino-olivary (Helweg’s tract) pathways. Spino-LRt neurons project ipsilaterally and contralaterally to the LRt in the medulla where they are involved in posture, reaching, and grasping (Alstermark and Ekerot, 2013; Jiang et al., 2015). Spino-olivary neurons reside in lamina V-VII of the spinal cord and project contralaterally to the inferior olive (IO) in the me- dulla (Oscarsson and Sjolund, 1977a,b; Berkley and Worden, 1978; Swenson and Castro, 1983a,b).

Furthermore, we find that cervical Atoh1-lineage neurons make the indirect spino-LRt and spino-olivary tracts rather than the direct lamina V-SCTs or dorsal horn SCTs as originally hypothesized and that thora- columbar Atoh1-lineage neurons project mainly locally within the spinal cord.

MiddletonStrick00a

Only DN, AIN project to thalamus:

KangJunBaekEtAl21

Even though DCN neurons in general project to a broad area, the overall projecting targets from each nucleus have different characteristics (Teune et al., 2000; Kebschull et al., 2020), such as more projections to the brainstem from the FN and more projections to the thalamus or midbrain from the IPN or DN. The anterior and posterior IPNs were also shown to have different projecting patterns (Teune et al., 2000; Lu et al., 2012). This suggests that each nucleus contributes to distinct functions. Detailed analyses further demonstrated different projection patterns from distinct groups of neurons in the same nucleus (Fujita et al., 2020; Kebschull et al., 2020).

Heterogeneity in the electrophysiological and anatomical properties has also been reported in DCN neurons (Czubayko et al., 2001; Aizenman et al., 2003; Sultan et al., 2003; Uusisaari and Knöpfel, 2012; Canto et al., 2016), although their functional relevance is not completely understood. A part of the heterogeneity arises from the three different types of projecting neurons (Baumel et al., 2009; Thanawalla et al., 2020). The majority of projecting neurons are glutamatergic neurons, which project to most of the extracerebellar target regions and send feedback signals to the cerebellar cortex (Houck and Person, 2015; Gao et al., 2016). Other well-known projecting neurons are GABAergic neurons, which specifically send feedback projections to the origin of the climbing fibers, i.e., the inferior olive (IO) (De Zeeuw et al., 1998), although a very recent study observed broad projections of GABAergic DCN neurons (Judd et al., 2021).

Various studies have shown the functional and structural relevance of the striped modular organization of the cerebellar cortex (De Zeeuw and Ten Brinke, 2015; Tsutsumi et al., 2015; Zhou J. et al., 2020; Beekhof et al., 2021). However, we will not consider the compartmental organization of the DCN in this article, because it is not yet clear how this organization is associated with the historically less well studied efferent pathways involved in non-motor cerebellar functions, which are the main topics of this article.

FujitaKodamaduLac20

Just the FN is crazy:

Big picture

The cerebellar vermis (from Latin vermis, "worm") is located in the medial, cortico-nuclear zone of the cerebellum, which is in the posterior fossa of the cranium.

Functionally, the vermis is associated with bodily posture and locomotion. The vermis is included within the spinocerebellum and receives somatic sensory input from the head and proximal body parts via ascending spinal pathways.[1]

WitterDeZeeuw15

AppsHawkes09

Much better than glickstein fig:

GlicksteinSultanVoogd11

Although there are no obvious morphological differences among different areas of the cerebellar cortex, analysis of its connections and chemoarchitecture, reveal that the Purkinje cells of the cortex are arranged in a series of precisely organized zones (Voogd, 1969) and that they differ in their histochemical properties (Hawkes and Leclerc, 1987).

Zones are present in the vermis, where they constitute a series of parasagittally-oriented stripes perpendicular to the fissures. The cerebellar hemispheres of mammals are typically made up of a continuous chain that departs from the midline in a succession of two loop-like structures (Bolk, 1906). A basic zonal organization is also present in the hemispheres, where the zones remain perpendicular to the cerebellar fissures. The functional nature of the zones is determined by the output of their target nuclei.

HenschkePakan20

Due to the indirect nature of cerebrocerebellar connections, it has been difficult to study their organization in precise detail. While a large body of research has mapped out the corticopontine projections (Brodal and Bjaalie, 1997; Glickstein, 1997; Leergaard and Bjaalie, 2007; Proville et al., 2014) or the pontocerebellar projections (Pijpers and Ruigrok, 2006; Pakan et al., 2010; Proville et al., 2014; Biswas et al., 2019), few studies have utilized neurotropic viruses (Kelly and Strick, 2003; Suzuki et al., 2012).

Therefore, the precise routes for information flow from the cortex to the cerebellum, including detailed terminal organization in the highly modular cerebellar cortex, have remained largely inferred. The integration of multimodal inputs to the cere- bellum is a fundamental operation that would allow for the precise coordination of sensory-driven movements within this brain region. Anatomically, the potential for a single granule cell to receive multimodal input via both descending motor cortex and ascending proprioceptive pathways has been shown in mice (Huang et al., 2013). However, the potential for co-innervation originating from various cortical areas spanning multiple modalities is unknown, even on the regional level in the cere- bellum, and has consequences for the role of cerebro-cerebellar-cerebro feedback loops in learning and predictive motor control (e.g. Chabrol et al., 2019).

Our results show that cerebrocerebellar information originating from different functional cortical regions shows significant spatial overlap within molecularly defined modules in crus I, PFL, and ver- mal lobules IV/V and VI. Therefore, all gross divisions of the cerebellar cortex (hemispheres, vestibu- locerebellum and vermis) have the potential for cortical multimodal influence, that is, the influence of the cerebral cortex is not strictly limited to the cerebellar hemispheres. These spatial overlaps constitute the fundamental anatomical basis for multimodal integration processes of cortical informa- tion at a modular and potentially cellular level (see also Proville et al., 2014). Functionally speaking, in vivo electrophysiological and imaging studies have shown that in rat (Ishikawa et al., 2015), mouse (Chabrol et al., 2015; Giovannucci et al., 2017; Markwalter et al., 2019) and mormyrid fish (Sawtell, 2010) single granule cells can respond to multiple modalities.

PisanoDhanerawalaKislinEtAl21

very broad patterns overall to thalamus -- need the equivalent map for CN!

BiswasLuoSarpongEtAl19

individual MF projections are rather broad!

[rcoreilly]

Complex Spikes

Hull20

However, such supervised learning also requires that CFs have access to a com- plete menu of erroneous actions in order to provide accurate error-based feedback. This may be practical only for a limited set of behaviors, and particularly for those where the environment can directly indicate what the correct action should have been. In other words, this supervised learning rule is ideal when there is a yoked relationship between stimulus and action, as is common for many well-studied cerebellar-dependent behaviors.

By responding via a hardwired neural pathway from the sensory periphery to the same corneal airpuff that produces a reflexive blink, CFs can accurately instruct a predictive eyelid closure accord- ing to the principles of supervised learning (Kim and Thompson, 1997; Medina et al., 2000).

In more complex motor behaviors without a fixed stimulus-action relationship, as well as many non-motor behaviors, it has been challenging to understand how such a supervised learning rule could provide an effective means for learning. In particular, in cases where the sensory information necessary for learning has no direct relationship to the movement that requires modification, or when the necessary sensory information is only applicable under a specific behavioral context, it is unclear whether or how CFs could generate such a supervised instructional signal.

For example, during arbitrary visuomotor reaching tasks, CF-driven Cspks in Purkinje cells have been shown to reflect predictive signals that are not consistent with motor errors (Kitazawa et al., 1998; Streng et al., 2017). Instead, these stud- ies have shown that Cspks can be predictive of task parameters such as reach destination, upcoming movement kinematics, or future posi- tion errors. Even during eyelid conditioning, recent evidence suggests that the cerebellum may be able to harness a wider range of distinct learning rules to modify behavior.

YES: this is key datapoint -- it is always doing sensory prediction!

(Ohmae and Medina, 2015): Importantly, the probability of Cspks on unexpected US trials was higher than for expected US trials where no conditioned response (CR) was generated, suggesting that differen- ces in sensory encoding of the US were not responsible for the differences in Cspk probability. These findings hence contradict what would be predicted by a supervised learning model, which would instead suggest the same Cspk probability on any trial type without a predictive eyelid closure to block the aversive corneal airpuff.

Another key feature of TD learning models is that teaching signals are modulated by experience to represent higher-order reinforcing stimuli. Again, CF activity met this criterion, as Cspks emerged in response to the conditioned stimulus (CS; i.e. an LED that reliably preceded the airpuff) after learning (Figure 2) (See also ten Brinke et al., 2015). No such CS-related Cspk responses would be predicted in supervised learning models.

Recent work has provided compelling evidence that the cerebellum may also contribute to reward- based reinforcement learning (Heffley and Hull, 2019; Heffley et al., 2018; Kostadinov et al., 2019; Larry et al., 2019). For example, two studies using calcium imaging to visualize Cspk activity in awake mice during operant learning tasks have now demonstrated that CFs can exhibit responses that are consistent with reward-based reinforcement learning signals (Heffley et al., 2018; Kostadinov et al., 2019; Figure 3). In each of these studies, mice were trained to execute a volun- tary action cued by a neutral sensory stimulus in order to receive reward. In these behaviors, both groups found that Cspks can reflect actions or events that predicted upcoming reward in a scalar manner that was proportional to reward expectation. In addition, these studies found that Cspks also report violated expectations by signaling when an expected reward is not delivered (Figure 3)

These results are not only consistent with reinforcement learning, but directly oppose the motor error hypothesis of CF activity. Specifically, because Cspk activity was generated in response to actions or events that accurately predicted upcoming reward (Heffley et al., 2018), this activity nec- essarily occurred when animals correctly executed movements rather than when movement was mis- executed.

Dude.. sensory prediction error!

To date, cerebellar-dependent sequence learning has been demonstrated only when movement feedback signals are reinforced by exogenously stimulated CF activity (Khilkevich et al., 2018). These data provide an exciting proof of principle for how cerebellar learn- ing can establish compound movements. However, it remains to be tested whether learned CF responses to conditioned stimuli can in fact support higher order conditioning and/or motor sequence learning.

It also remains plausible that learned CF responses to conditioned stimuli can allow further modification of movement by other means, for example by directly modulating the activity of CbN neurons (Ten Brinke et al., 2019; Ten Brinke et al., 2017), or may serve a different purpose altogether by enabling cerebellar output to downstream brain regions. Testing such predic- tions will be crucial for understanding how learned, conditioned stimulus-driven CF signals are har- nessed for modifying behavior.

key point: signed CS for eyeblink, not for reward association learning..

In particular, neither Heffley et al. nor Kostadinov et al. observed a decrease in Cspk activity when expected reward was not delivered. Instead, both studies reported elevated Cspk activity in response to defied expectations. Such responses are consistent with unsigned pre- diction errors, but not with the types of signed prediction errors typically associated with TD learning (Figure 4). The decreases in Cspk activity demonstrated by Ohmae et al. are ideal to promote extinction learning during eyeblink conditioning when the aversive conditioned stimulus is absent (Medina et al., 2002), and may therefore also be a hallmark of behaviors with fixed stimulus-action relationships. In contrast, the increased Cspk responses shown by Heffley et al. and Kostadinov et al. may promote exploration and new associations during behavior that requires flexible stimulus-action relationships. Hence, despite the commonality of learned Cspk responses to higher-order reinforcing stimuli across studies, it remains to be determined how specific task requirements and location within the cerebellum determine the CF response to violated expectations.

Origins of climbing fiber reinforcement learning signals

For example, in eyelid conditioning, the cor- neal airpuff used as a US is transmitted to the IO via the trigeminal ganglion (Kubo et al., 2018; Swenson and Castro, 1983; Van Ham and Yeo, 1992), reliably triggering CF input to the cerebel- lum. This ‘hardwired’ pathway allows CFs to respond with high fidelity in a manner consistent with supervised learning based on the causal association between stimulus and movement. However, the IO also receives indirect input from both cortical and subcortical brain regions that may allow it to generate more complex teaching signals than can be produced by salient sensory input from the environment (Ten Brinke et al., 2019).

One such key source of input to the IO is the mesodiencephalic junction (MDJ) (De Zeeuw et al., 1998). This region of the midbrain is composed of multiple nuclei, some of which integrate cerebel- lar CbN output and project to either downstream neurons in the spinal cord (Keifer and Houk, 1994) or the IO (Onodera, 1984). Because cerebellar learning has been shown to produce new mossy fiber collaterals in the CbN (Boele et al., 2013; Kleim et al., 2002; Lee et al., 2015; Weeks et al., 2007), it has been speculated that such collaterals may convey information about learned conditioned stimuli to the MDJ, and in turn to the IO (Ten Brinke et al., 2019; Ten Brinke et al., 2017). In this model, mossy fiber pathways carrying learned information about conditioned stimuli could be indirectly translated to the IO via these new collaterals, producing instructive rein- forcement learning signals in CFs. Notably, such an extended polysynaptic pathway may partly explain why CF signals related to conditioned stimuli tend to have longer latencies than the US- related CF signals that are translated more directly from the sensory periphery.

The MDJ also integrates descending input from sources that include cortical pyramidal tract neu- rons (Veazey and Severin, 1982). Such input likely allows the IO to represent higher order cortical computations, and is therefore a good candidate to transmit the types of reward prediction and TD- learning signals that have been shown recently.

The IO also receives inhibitory input, and contains some local interneurons, raising the possibility that local computations may also enrich the repertoire of CF responses (De Zeeuw et al., 1998). For example, the IO receives inhibitory input from the cerebellar nuclei (CbN) (Bengtsson and Hesslow, 2006). These inhibitory projections participate in extinction learning (Medina et al., 2002), but could also play a role in gating CF activity or sculpting responses to incoming excitation. For example, inhibitory inputs could decouple stimulus-action relationships resulting from peripheral IO inputs, enabling other pathways to dominate when learning requires information from other sources. Indeed, multiple studies have found that CF activity exhibits context dependence, and that sensory responses can be actively suppressed under certain behavioral conditions (Apps, 1999; Apps and Lee, 1999; Gellman et al., 1985; Horn et al., 1996; Kim et al., 1987).

Completely missing the idea that this is the predictive input.

Whether inhibitory CbN pro- jections play a role in such context-dependent IO suppression remains unclear. More broadly, it has remained challenging to establish any clear predictions about the influence of discrete pathways in generating CF learning signals, as there remains an incomplete description of inputs to the IO, as well as a limited understanding of when different pathways are active and how the IO integrates input to generate CF responses in vivo. Thus, a key step in understanding how CFs can produce complex teaching signals such as those necessary for reinforcement learning will be to establish a more detailed map of input to the IO, and to measure the behavioral contexts under which specific input pathways recruit CF activity.

In contrast with classic models, recent work has challenged the idea that granule cells generate sparse representations, and shown that they can instead exhibit dense and redundant responses during several behaviors (Giovannucci et al., 2017; Knogler et al., 2017; Ozden et al., 2012; Sylvester et al., 2017). Such results are surprising, and may suggest that it is necessary to rethink how the cerebellum forms unique sensorimotor associations. However, an alternative possibility is that some aspects of the original Marr-Albus models require revision. For example, pattern separa- tion need not require sparse coding (Cayco-Gajic and Silver, 2019). In particular, as argued by Cayco-Gajic and Silver, pattern separation can be achieved without sparse coding if inputs are still expanded and decorrelated, allowing dense granule cell responses to effectively discriminate com- plex, high-dimensional inputs.

Indeed, recent calcium imaging data further supports the idea that granule cells can represent efference copy information (Giovannucci et al., 2017). Using calcium imaging to measure the responses of granule cells across eyelid conditioning, Giovannucci et al. revealed learned representations of the conditioned eyelid closure that can match, or even pre- cede the eyelid movement after learning.

Several elegant studies have now begun to reveal that the sodium-based action potentials of Purkinje cells (so-called ‘simple spikes’) can establish complex kinematic predictions about movement across diverse behaviors (Brown and Raman, 2018; Chen et al., 2016; Ebner and Pasalar, 2008; Herzfeld et al., 2015; Pasalar et al., 2006). Such data support the notion that the cerebellum can harnesses a type of predictive computation often termed a ‘forward model’. A forward model can be most simply conceptualized as an estima- tion of the immediate future based on a copy of the current motor command and current sensory information.

A key virtue of such predictions is that they can be used as a more rapid substitute for external feedback (e.g. the sensory consequences of movement) to enable anticipatory actions. Such forward model predictions can also be compared with subsequent sensorimotor feedback to assess differences between expectation and outcome. When there is a mismatch, termed a ‘sensory predic- tion error’, the forward model can be updated via learning mechanisms (e.g. synaptic plasticity) in accordance with current conditions.

Ya think?

In cases where CF activity is strongly linked to learning, the primary model suggests that these signals serve to instruct long-term synaptic depression (LTD) of granule cell synapses onto Purkinje cells (Albus, 1971; Ito, 1972). This mechanism is appealing because Purkinje cells are inhibitory, and exhibit high convergence onto target neurons in the CbN that form the output of the cerebellum (Person and Raman, 2012). Thus, predictive cerebellar output from CbN neurons would be greatly facilitated by an appropriately timed disinhibition that could be achieved by reducing the simple spiking of Purkinje cells (Heiney et al., 2014; Figure 6).

Indeed, while current evidence suggests a role for plasticity of various types and at multiple sites in the cerebellar circuit (Boyden et al., 2004; Carey and Lisberger, 2002; Gao et al., 2012), there remains considerable evidence that CF-driven LTD, or at least CF-driven reductions in PC simple spiking, is involved with many forms of cerebellar learning (Ito et al., 2014). In particular, studies of learned eye movements have provided the most direct evidence to date for a causal link between Cspk-driven reductions in Purkinje cell simple spiking and learning (Herzfeld et al., 2018; Kimpo et al., 2014; Medina and Lisberger, 2008; Yang and Lisberger, 2014; Figure 7). This work has shown that Cspks are highly correlated with a depression of PC simple spiking and learning on a single trial basis. Moreover, these studies have made a key link between the duration of Cspks and learning, revealing that these signals are graded, likely due to graded presynaptic CF activity (Gaffield et al., 2019), in a manner that scales with the amount of single trial learning.

Despite key efforts in this direction (Schonewille et al., 2011; Yamaguchi et al., 2016), conclusive tests of how LTD at PC synapses contributes to learning will require manipulations that are not only cell-type specific, but also temporally specific in order to overcome the complications of circuit compensation that are inherent to chronic genetic strategies.

For example, during eyelid conditioning, CbN neurons become active to the conditioned stimulus to enable a predictive eyelid closure by activating downstream motor neurons (McCormick and Thompson, 1984).

Key ref: Brooks et al 2015:

By recording from the CbN of awake behaving primates, Brooks and colleagues showed that these neurons can dynamically track the difference between expected motor output and sensory feedback, consistent with the computation of sensory prediction error during voluntary head move- ments (Brooks et al., 2015). In particular, they found that CbN neurons responded to mismatches between expected and actual head movements on a trial-by-trial basis (Figure 8). Moreover, these prediction-dependent CbN responses were continuously updated with new learning.

Notably, CbN neurons in this study were modulated in the same direction regardless of the direction of sensorimotor mismatch. Specifically, CbN neurons elevated their firing regardless of whether head movement was unexpectedly restricted, or unexpectedly released from restriction during extinction learning. Such unidirectional signaling of mismatch is likely appropriate to drive stabilizing vestibluo-spinal reflexes (Roy and Cullen, 2001; Roy and Cullen, 2004) and to ensure stable perception that accounts for self-motion (Dale and Cullen, 2019). It has remained challenging, however, to establish such causal relationships between CbN activity and behavior, as tools to selectively manipulate CbN neurons in a manner that accounts for ongoing behavior have not been available until recently.

Becker & Person 2019:

To begin addressing the casual relationship between CbN activity and behavior, a recent study has used closed-loop optogenetic manipulations to alter the firing of CbN neurons during move- ment (Becker and Person, 2019). By both activating and inhibiting CbN neurons at different time- points during skilled reaching, this study revealed that the activity of these cells contributes to predictively controlling the endpoint of reaches in real time. Specifically, the authors found that CbN output contributed unidirectionally to limb movement, adding velocity toward the body regardless of limb movement direction. These data suggest that the cerebellum can contribute at least part of the motor command necessary to regulate ongoing movements on a millisecond timescale, calibrat- ing behavior even during steady state performance after learning has occurred. Notably, such CbN activity that persists after learning differs considerably from the patterns measured by Brooks and colleagues, where CbN output was abolished after learning (provided movement proceeded as expected). In absence of a clear distinction between the features of learning across studies such as these, it seems that we are only at the early stages of understanding how cerebellar output 1) con- tributes to motor control across diverse behaviors and learning paradigms, and 2) conforms to pre- dictions about the implementation of cerebellar forward models.

Thus, while forward model explanations have gained considerable support from behavioral studies in humans (Izawa et al., 2012; Morton and Bastian, 2006) and animals (Machado et al., 2015; Pasalar et al., 2006), it will be necessary to further probe the links between theory and newly emerging datasets. For example, it will be crucial to determine whether and how behavioral demands, cerebellar region, and down- stream target area dictate the mode of cerebellar output. Such efforts will be challenged by the wid- ening range of brain regions and behaviors and that the cerebellum contributes to, including those that are now recognized to involve complex cortical computations.

Training thalamus:

(Chabrol et al., 2019; Gao et al., 2018). In addition to descending rubrospinal pathways, the cerebellum is heavily connected to the neocortex disynaptically via the thalamus... Remarkably, by optogenetically inhibiting cerebellar CbN neu- rons, Gao et al. found that cerebellar output was necessary for ramping activity in ALM during a motor discrimination task (Figure 9). Moreover, disrupting cerebellar CbN output impaired motor- based decisions by introducing a motor bias, but did not disrupt motor output per se

Together, these ALM results are consistent with studies in non-human primates showing that neurons in the lateral CbN can exhibit predictive ramping responses during a delay period prior to movement (Ashmore and Sommer, 2013; Ohmae et al., 2017). While such CbN ramping responses have been implicated in action timing, it is reasonable to extrapolate that temporal sig- nals of this type could be used by the neocortex for a variety of timing computations. Similar to delay-interval ramping, CbN neurons have also been shown to predictively represent the timing of periodic stimuli with oscillatory responses that increase prior to each stimulus presentation. Such responses appear to reflect temporal predictions rather than expectations about other stimulus features, as the same neurons preferentially sig- naled when temporal periodicity was unexpect- edly violated (Kameda et al., 2019; Ohmae et al., 2013). Thus, it appears that the cerebellum can transmit various timing predic- tions to the neocortex, an idea that is borne out by the impairment of movement timing when the

Notably, however, the cerebellar thalamocortical pathway projecting to primary motor cortex appears not to signal via anticipatory pre-movement ramping. Rather, motor cortical responses driven by this pathway exhibit transient excitation followed by a long-lasting recruitment of cortical inhibition (Nashef et al., 2018). While it is unclear how the CbN neurons projecting to the motor thalamocortical pathway fired in that study, the data are at least suggestive that predictive ramping activity is not the sole mode of cerebellar output to motor cortical areas during timing tasks.

[rcoreilly]

Corrections as training signal

Verduzco-FloresOReilly15 proposed that learning could be based on applying a later corrective motion in an anticipatory way, to avoid the error in the first place. this is related to Fujita’s feed-forward associative learning model (Fujita, 2005).

It kinda leaves the question of how the corrective action is taken in the first place, but presumably sensory information is available to show the error, and the corrective action can operate on that, in the same way it generates the initial action toward a target.

In any case, ShadmehrSmithKrakauer10 address this and show that it is not supported across 3 studies:

One way to think of motor learning is to imagine that the motor commands that corrected the movement might be added with a slight time advance to the motor com- mands that initiate the next movement. The er- ror feedback learning theory of the cerebellum relies on this motor correction (Kawato 1996).

To address this question, experiments have attempted to dissociate the effect of error- driven motor corrections from error signals themselves. In saccade adaptation experiments this has been done by presenting endpoint er- rors (sensory prediction errors) under condi- tions that reduce motor corrections (Wallman & Fuchs 1998, Noto & Robinson 2001). In reach adaptation paradigms this has been done by making a rapid movement so that there is a reduced possibility of correcting that move- ment (Tseng et al. 2007). All three experiments demonstrate that adaptation of the motor com- mands is driven directly by error signals rather than error corrections.

Looks like Kawato 1996 was the better ref.