Skip to content
Robotic Brain Universe

The cerebellum and motor learning

How the 'little brain', with more neurons than the cerebral cortex, predicts and corrects movement, and how it shapes thinking and AI models of learning.

Intermediate · about 9 min · updated 2026-10-02 · awaiting clinical review

See it in 3D:

Illustrative simulation excitatory inhibitory

The cerebellar circuit of mossy, parallel and climbing fibres and Purkinje cells; Marr's and Albus's theories and climbing-fibre long-term depression; internal models, force-field and prism adaptation; fast and slow learning; ataxias and the cerebellar cognitive affective syndrome; error-driven learning, forward models and the two-state model; robotic and imaging tools; and cerebellum-inspired models of learning.

Contents
  1. The little brain that does a lot
  2. What the cerebellum is
  3. Why we need a cerebellum
  4. How the cerebellum learns
  5. When learning happens: fast and slow
  6. When the cerebellum fails
  7. The mathematics of motor learning
  8. Technology: robots, scanners and models
  9. Milestones
  10. Frontiers
  11. Check yourself

The little brain that does a lot

The cerebellum, tucked under the back of the cerebrum, is smaller than the cerebral cortex yet across 19 mammal species it holds on average 3.6 neurons for every neuron in the cerebral cortex. Its surface is so tightly folded that, flattened out, the human cerebellar cortex would form a strip about 10 cm wide and almost a metre long, 78% of the area of the neocortex.[1,2]

It is the brain's error-correction engine. When you learn to throw while wearing prism glasses, reach in a new force field or simply keep a movement smooth, the cerebellum compares what you intended with what happened and adjusts. Theories of how it does this, proposed by David Marr in 1969 and James Albus in 1971, treated it as a learning machine.[3,4,5,6]

It is not only about movement. Damage to the posterior cerebellum can change planning, language, spatial thinking and personality, and most of the human cerebellum is connected with association areas of the cerebral cortex rather than motor areas.[7,8]

What the cerebellum is

The cerebellar cortex has a strikingly regular circuit. Mossy fibres carry information about context to granule cells, whose parallel fibres contact the Purkinje cells; each Purkinje cell also receives a single powerful climbing fibre from the inferior olive. The Purkinje cells are the only output of the cerebellar cortex.[5,9]

The cerebellum connects with the rest of the brain through three pairs of peduncles; the middle peduncle carries input from the pontine nuclei, the inferior peduncle carries climbing fibres from the inferior olive, and the superior peduncle carries output towards the thalamus.[3]

Key numbers

Cerebellar neurons per cerebral cortical neuron, average across 19 mammal species
3.6[1]
Surface area of the human cerebellar cortex (shrinkage-corrected)
1,590 cm²[2]
As a share of the neocortex's surface area: human, macaque
78%, about 33%[2]
Known dominant spinocerebellar ataxia genes caused by repeat expansions (2010)
11 of 18[10]

Why we need a cerebellum

Movements have to be planned around delays: sensory feedback arrives too late to steer a fast movement. Wolpert, Miall and Kawato reviewed evidence that the cerebellum contains internal models of the body: inverse models that work out the commands needed for a desired movement, and forward models that predict the consequences of a command so the brain need not wait for feedback.[11]

The cerebellum also keeps movements calibrated as the body and world change. People wearing prism glasses initially throw in the direction the prisms bend their gaze, then adapt to hit the target; patients with lesions of the olive–cerebellar system adapt poorly or not at all.[3]

How the cerebellum learns

Two theories. Marr proposed that each Purkinje cell learns to recognise, from its mossy-fibre input, the context in which its climbing fibre fired, so that a learned action can later run without instruction. Albus proposed that the cerebellum works like a perceptron trained by a teacher, with the climbing fibre acting as an error signal that weakens the synapses that were active.[5,6]

Long-term depression. Ito and colleagues found that pairing vestibular mossy-fibre input with climbing-fibre stimulation depressed the Purkinje cells' responses to that input, with a slow depression lasting for an hour, while inputs not paired with the climbing fibre were unaffected: a synaptic learning rule driven by the climbing-fibre teaching signal.[9]

Learning a model of the world. When people reached while a robot applied a force field to their hand, their paths were at first badly distorted and then straightened with practice. Switching the field off produced roughly mirror-image after-effects, suggesting that the brain had gradually built a model of the field.[4]

The cerebellar learning circuitMossy fibrescontext: where the body is, whatis plannedGranule cellsparallel fibresPurkinje cellsthe only output of cerebellarcortexInferior oliveclimbing fibre: errorDeep cerebellar nucleito thalamus and brainstemCorrected movementsmoother, better calibratedplastic synapsesteaching signalerrors
The cerebellar learning circuit. Context arrives on mossy and parallel fibres; when something goes wrong, the climbing fibre from the inferior olive signals the error and weakens the parallel-fibre synapses that were active, so the Purkinje cell's output, and the movement, is corrected next time.[3,5,6,9]
Text version of the diagram
  1. Mossy fibres: context: where the body is, what is planned. Leads to Granule cells.
  2. Granule cells: parallel fibres. Leads to Purkinje cells (plastic synapses).
  3. Purkinje cells: the only output of cerebellar cortex. Leads to Deep cerebellar nuclei.
  4. Inferior olive: climbing fibre: error. Leads to Purkinje cells (teaching signal).
  5. Deep cerebellar nuclei: to thalamus and brainstem. Leads to Corrected movement.
  6. Corrected movement: smoother, better calibrated. Leads to Inferior olive (errors).

When learning happens: fast and slow

Motor adaptation is faster than once thought. Smith, Ghazizadeh and Shadmehr showed that within minutes two processes are at work: one responds strongly to error but forgets quickly, the other responds weakly but retains well.[12]

Their interaction predicts surprising effects. After adapting to a perturbation and then quickly unlearning it, people show spontaneous recovery of the earlier adaptation when error feedback is removed, and the same two-state model explains savings (faster relearning), interference and rapid unlearning.[12]

When the cerebellum fails

Losing adaptation. In prism-throwing experiments, patients with infarcts in the territory of the posterior inferior cerebellar artery usually lost the ability to adapt while showing little or no ataxia, whereas damage in the superior cerebellar artery territory or the cerebellar thalamus usually caused ataxia but left adaptation intact.[3]

Inherited ataxias. The dominantly inherited spinocerebellar ataxias are rare, and many involve more than the cerebellum. By 2010, 11 of the 18 known genes acted through repeat expansions; others affect glutamate and calcium signalling, ion channels or mitochondria.[10]

Thinking and feeling. In 20 patients with disease confined to the cerebellum, lesions of the posterior lobe and vermis caused impairments of planning, verbal fluency, abstract reasoning and working memory, difficulties with spatial cognition, personality change and language problems: the cerebellar cognitive affective syndrome. Anterior lobe lesions produced only minor changes.[7]

The mathematics of motor learning

The cerebellum is often described as a supervised learner that builds predictive models, which makes its mathematics close to that of machine learning.[6,11]

Error-driven weight change at parallel-fibre synapses[6,9]
Δwi=− η  e  xi\Delta w_i = -\,\eta\; e\; x_i

When the climbing fibre signals an error ee, the synapses from parallel fibres that were active (xi>0x_i > 0) are weakened in proportion to their activity, the kind of supervised rule Albus proposed and Ito's long-term depression implements. It has the same form as the delta rule of machine learning.

Symbols in Error-driven weight change at parallel-fibre synapses
SymbolMeaningUnit
wiw_istrength of the synapse from parallel fibre i—
xix_iactivity of parallel fibre i—
eeerror signalled by the climbing fibre—
η\etalearning rate—
Forward model[11]
x^t+1=f ⁣(xt, ut),et+1=xt+1−x^t+1\hat{\mathbf{x}}_{t+1} = f\!\left(\mathbf{x}_t,\, \mathbf{u}_t\right), \qquad \mathbf{e}_{t+1} = \mathbf{x}_{t+1} - \hat{\mathbf{x}}_{t+1}

A forward model predicts the next state of the body from its current state and the motor command, before feedback arrives; the difference between prediction and outcome is the error used to improve the model.

Symbols in Forward model
SymbolMeaningUnit
xt\mathbf{x}_tstate of the body (positions, velocities) at time t—
ut\mathbf{u}_tmotor command—
x^t+1\hat{\mathbf{x}}_{t+1}predicted next state—
et+1\mathbf{e}_{t+1}sensory prediction error—
Two-state model of adaptation[12]
xf(n+1)=Af xf(n)+Bf e(n),xs(n+1)=As xs(n)+Bs e(n),x=xf+xsx_f(n{+}1) = A_f\,x_f(n) + B_f\,e(n), \quad x_s(n{+}1) = A_s\,x_s(n) + B_s\,e(n), \quad x = x_f + x_s

Adaptation is the sum of a fast process that learns quickly (Bf>BsB_f > B_s) but forgets quickly (Af<AsA_f < A_s) and a slow process that learns slowly but retains well. The interaction of the two reproduces savings, interference, spontaneous recovery and rapid unlearning.

Symbols in Two-state model of adaptation
SymbolMeaningUnit
xf,xsx_f, x_sfast and slow states of adaptation—
Af,AsA_f, A_sretention factors (between 0 and 1)—
Bf,BsB_f, B_slearning rates from error—
e(n)e(n)error on trial n—

Technology: robots, scanners and models

Robotic manipulanda. A robot that applies controlled forces to the hand made it possible to study how people learn new dynamics, and to reveal the internal model through after-effects when the forces are switched off.[4]

Mapping the cerebellum. Resting-state functional MRI in 1,000 people showed an inverted body map in the anterior lobe and an upright one in the posterior lobe, and that most of the human cerebellum maps to association areas of the cerebral cortex; primary visual cortex is not represented.[8]

Unfolding it. Reconstructing every individual folium from high-resolution post-mortem MRI gave the first full surface of the human cerebellar cortex, 1,590 cm², against about a third of the neocortex's area in the macaque.[2]

Cerebellum-inspired AI. Inspired by deep-learning methods, Boven and colleagues modelled the cerebellum as a network that predicts feedback for the cerebral cortex, decoupling cortical learning from slow feedback; the model learned faster and showed fewer dysmetria-like errors, in motor and cognitive tasks.[13]

Milestones

Understanding the cerebellum

  1. 1969Marr's theory: the cerebellum learns motor skills.[5]
  2. 1971Albus models the cerebellum as a perceptron trained by climbing-fibre errors.[6]
  3. 1982Climbing-fibre-induced long-term depression is shown in Purkinje cells.[9]
  4. 1994People learn internal models of novel force fields.[4]
  5. 1996Olivocerebellar lesions impair prism adaptation.[3]
  6. 1998The cerebellar cognitive affective syndrome is described; internal models in the cerebellum are reviewed.[7,11]
  7. 2006A two-state model explains fast and slow motor adaptation.[12]
  8. 2010Cortical and cerebellar neuron numbers are shown to scale together across mammals.[1]
  9. 2011Most of the human cerebellum is mapped to association networks.[8]
  10. 2020The human cerebellar surface is reconstructed: 78% of the neocortex.[2]
  11. 2023A cerebro-cerebellar model links the cerebellum to feedback prediction in deep learning.[13]

Frontiers

The human cerebellum's surface area is 78% of the neocortex's, against about 33% in the macaque, suggesting a prominent role in the evolution of distinctively human behaviour and cognition.[2]

Computational models now treat the cerebellum as a general prediction machine for the whole brain, supplying the cortex with predicted feedback so that it can learn when real feedback is sparse or late.[13]

Check yourself

Check yourself

  1. Roughly how many cerebellar neurons are there for every cerebral cortical neuron across mammals?
    Show answer

    About 3.6.

  2. Which cell is the only output of the cerebellar cortex, and which input carries the error signal?
    Show answer

    The Purkinje cell; the climbing fibre from the inferior olive.

  3. What did Ito and colleagues show?
    Show answer

    Pairing mossy-fibre input with climbing-fibre stimulation depresses Purkinje cell responses to that input (long-term depression).

  4. What are forward and inverse internal models?
    Show answer

    A forward model predicts the consequences of a command; an inverse model computes the command needed for a desired movement.

  5. What did the force-field after-effects reveal?
    Show answer

    That people had learned a model of the force field, producing mirror-image errors when it was removed.

  6. What is the cerebellar cognitive affective syndrome?
    Show answer

    Executive, spatial, personality and language changes after posterior cerebellar and vermis lesions.

  7. What does the two-state model of adaptation explain?
    Show answer

    Savings, interference, spontaneous recovery and rapid unlearning, from a fast and a slow learning process.

Glossary[5,9,10,11,12,13]

Cerebellum
The 'little brain' at the back of the skull that calibrates movement and supports cognition.
Purkinje cell
The large output neuron of the cerebellar cortex.
Climbing fibre
An input from the inferior olive that signals errors to a Purkinje cell.
Mossy and parallel fibres
Inputs carrying context to granule cells, whose parallel fibres contact Purkinje cells.
Long-term depression
A lasting weakening of synapses, here of parallel-fibre inputs after climbing-fibre activity.
Internal model
A neural representation that predicts or inverts the behaviour of the body or a tool.
Motor adaptation
Recalibrating movements to a changed body or environment.
Ataxia
Loss of coordination of movement.
Dysmetria
Over- or undershooting the intended target of a movement.
Spontaneous recovery
Re-emergence of an adaptation that had apparently been unlearned.

References

  1. Herculano-Houzel S. Coordinated scaling of cortical and cerebellar numbers of neurons. Frontiers in Neuroanatomy 2010;4:12. doi:10.3389/fnana.2010.00012
  2. Sereno MI, Diedrichsen J, Tachrount M, Testa-Silva G, d'Arceuil H, De Zeeuw C. The human cerebellum has almost 80% of the surface area of the neocortex. Proceedings of the National Academy of Sciences of the USA 2020;117(32):19538-19543. doi:10.1073/pnas.2002896117
  3. Martin TA, Keating JG, Goodkin HP, Bastian AJ, Thach WT. Throwing while looking through prisms: I. Focal olivocerebellar lesions impair adaptation. Brain 1996;119(4):1183-1198. doi:10.1093/brain/119.4.1183
  4. Shadmehr R, Mussa-Ivaldi FA. Adaptive representation of dynamics during learning of a motor task. The Journal of Neuroscience 1994;14(5):3208-3224. doi:10.1523/JNEUROSCI.14-05-03208.1994
  5. Marr D. A theory of cerebellar cortex. The Journal of Physiology 1969;202(2):437-470. doi:10.1113/jphysiol.1969.sp008820
  6. Albus JS. A theory of cerebellar function. Mathematical Biosciences 1971;10(1-2):25-61. doi:10.1016/0025-5564(71)90051-4
  7. Schmahmann JD, Sherman JC. The cerebellar cognitive affective syndrome. Brain 1998;121(4):561-579. doi:10.1093/brain/121.4.561
  8. Buckner RL, Krienen FM, Castellanos A, Diaz JC, Yeo BTT. The organization of the human cerebellum estimated by intrinsic functional connectivity. Journal of Neurophysiology 2011;106(5):2322-2345. doi:10.1152/jn.00339.2011
  9. Ito M, Sakurai M, Tongroach P. Climbing fibre induced depression of both mossy fibre responsiveness and glutamate sensitivity of cerebellar Purkinje cells. The Journal of Physiology 1982;324(1):113-134. doi:10.1113/jphysiol.1982.sp014103
  10. Durr A. Autosomal dominant cerebellar ataxias: polyglutamine expansions and beyond. The Lancet Neurology 2010;9(9):885-894. doi:10.1016/S1474-4422(10)70183-6
  11. Wolpert DM, Miall RC, Kawato M. Internal models in the cerebellum. Trends in Cognitive Sciences 1998;2(9):338-347. doi:10.1016/S1364-6613(98)01221-2
  12. Smith MA, Ghazizadeh A, Shadmehr R. Interacting adaptive processes with different timescales underlie short-term motor learning. PLoS Biology 2006;4(6):e179. doi:10.1371/journal.pbio.0040179
  13. Boven E, Pemberton J, Chadderton P, Apps R, Costa RP. Cerebro-cerebellar networks facilitate learning through feedback decoupling. Nature Communications 2023;14:51. doi:10.1038/s41467-022-35658-8

Related readings

  • The motor system

    How the brain turns intention into movement, from the redrawn motor homunculus to the laws of reaching and the technology that restores it.

    Intermediate

  • Neural networks, biological and artificial

    How brains and machines learn from experience, the maths both share, and why AI and neuroscience keep borrowing from each other.

    Intermediate

  • The basal ganglia and reward

    How deep brain circuits choose actions, build habits and learn from reward, and why the dopamine signal inspired a branch of AI.

    Intermediate

  • The brainstem and cranial nerves

    How the brainstem keeps us breathing and awake, what the twelve cranial nerves do, how they are tested, and what happens when they fail.

    Intermediate

Template anatomy for education. Not patient-specific. Not for clinical decision-making.