A theory of cerebellar cortex

TL;DR — In 1969 David Marr read the cerebellum's anatomy as an algorithm: mossy fibers carry a dense "context" vector, ~50 billion granule cells expand it into a sparse high-dimensional code, and each Purkinje cell is a one-layer supervised readout whose parallel-fiber synapses are trained by a climbing-fiber teaching signal. This is expansion recoding + random features + a perceptron with a labeled error line, written down before that vocabulary existed and derived almost entirely from cell counts and wiring. The plasticity prediction was later confirmed (as LTD, with the sign flipped), the theory of sparse expansion has since been made rigorous, and the architecture is exactly what robot foundation models are missing today: a fast, data-efficient, online adapter next to a slow pretrained policy.

The idea Marr had

The cerebellum is anatomically strange in a way that begs for a computational reading. Its cortex is a crystalline, repeated circuit. Mossy fibers bring input from everywhere — spinal cord, vestibular system, cerebral cortex. Each granule cell samples only about 4 mossy fibers, and there are an enormous number of granule cells — the majority of all neurons in the brain live in this one layer. Their axons rise and split into parallel fibers, which run through the flat, fan-like dendritic trees of Purkinje cells. Each Purkinje cell receives on the order of 10^5 parallel-fiber synapses — and exactly one climbing fiber from the inferior olive, which wraps around it and fires it unconditionally.

Marr's reading: the granule layer performs expansion recoding. A dense, correlated mossy-fiber pattern is re-expressed as a sparse pattern over a vastly larger population (his "codons" — small subsets of active mossy fibers detected by individual granule cells). Golgi cell inhibition acts as a thermostat keeping the granule code sparse regardless of input density. Sparseness plus dimensionality expansion makes patterns that were tangled in the input nearly orthogonal in the granule code — so a linear classifier can now separate them.

The linear classifier is the Purkinje cell. The climbing fiber is the label. Whenever the olive fires — in Marr's story, signaling a cerebral "instruction" for an elemental movement — the conjunction of climbing-fiber activity and parallel-fiber activity modifies the active parallel-fiber synapses. After learning, the mossy-fiber context alone drives the Purkinje cell, and the movement runs without cerebral supervision. Marr predicted: (a) parallel-fiber → Purkinje synapses are modifiable by this conjunction, (b) nothing else in the cerebellum is, (c) Golgi cells regulate granule sparseness. He even ran capacity arguments — how many contexts a Purkinje cell can store — from synapse counts.

Why it could not work in 1969

No experimental access to plasticity. Marr's central prediction was about synaptic modification under a specific conjunction of inputs. In 1969 nobody had demonstrated activity-dependent synaptic plasticity anywhere in the mammalian brain — Bliss and Lømo's hippocampal LTP paper is from 1973. Testing the parallel-fiber prediction required simultaneously driving parallel fibers and climbing fibers while recording one Purkinje cell's synaptic efficacy — technically out of reach for over a decade.

No connectome, only statistics. Marr's argument leaned on wiring statistics from Eccles, Ito, and Szentágothai's 1967 anatomy. Whether granule connectivity was actually random (his implicit assumption) was unmeasurable; electron-microscopy reconstruction of even a cubic millimeter was science fiction.

No theoretical frame to test the math. The perceptron existed (Rosenblatt 1958, Cover's 1965 capacity result), but there was no theory of why sparse random expansion helps, no notion of random-feature kernels, no way to simulate the claims. A machine of that era ran at roughly 10^5–10^6 operations per second with a few hundred kilobytes of memory; a single Purkinje cell's 10^5 synapses updated over thousands of trials was a serious computation, and a granule layer with realistic expansion was untouchable. Today one GPU delivers roughly 10^{14}–10^{15} ops/s — nine to ten orders of magnitude — and the whole model is an afternoon's notebook.

What changed

The plasticity prediction was confirmed — with the sign flipped. Albus (1971) independently derived essentially the same theory but argued the conjunction should depress parallel-fiber synapses (Purkinje cells are inhibitory onto their targets, so weakening is the natural "correction"). Ito and colleagues demonstrated exactly this long-term depression in 1982. Marr–Albus–Ito is now the textbook account, one of the few cases where a circuit-level learning theory made a specific synaptic prediction that experiment later verified.

The math got made rigorous. Random-feature theory (Rahimi & Recht, 2007) explains why a fixed random nonlinear expansion plus a trained linear readout approximates kernel regression. Babadi & Sompolinsky (2014) and Litwin-Kumar, Harris, Axel, Litwin-Kumar & Abbott (2017) analyzed cerebellum-like sparse expansion directly, showing that sparse connectivity of roughly four inputs per granule cell is near-optimal for the dimensionality of the representation given wiring cost — turning Marr's anatomical curiosity into a derived optimum.

Convergent evolution vindicated the architecture. The Drosophila mushroom body (sparse Kenyon-cell expansion + dopaminergic teaching signal onto a small readout) and the cerebellum-like electrosensory lobes of electric fish (Bell's adaptive-filter work) implement the same motif. When evolution builds the same three-layer supervised machine repeatedly, the design is probably right.

Connectomics arrived — and complicated the story productively. Dense EM reconstruction of mouse cerebellar tissue (e.g., Nguyen et al., Nature 2023, to my knowledge) found mossy-to-granule wiring is more redundant and structured than the random model assumes — apparently trading some dimensionality for reliability. Marr's assumption is now a measurable, and measurably wrong-in-detail, quantity. Likewise, climbing fibers turn out to carry graded, probabilistic error information rather than a binary event, plasticity exists at many cerebellar sites beyond the parallel-fiber synapse, and mice with impaired LTD can still show some motor learning (Schonewille et al., 2011 — worth reading as the honest counterpoint). The core architecture stands; the 1969 simplifications do not.

A 2026 revival: the cerebellar sidecar

Here is the experiment I would actually run. Modern robot foundation policies (VLA models like OpenVLA or π0-style systems) are the "cerebral cortex" of the analogy: slow, general, expensive to update, running at maybe 5–50 Hz, and systematically wrong about this particular robot's miscalibrated wrist, worn gripper, or payload of the day. The cerebellum's job, on Marr's reading and on the modern internal-model reading (Wolpert, Kawato), is precisely to learn fast context-dependent corrections beside a slower controller.

The sidecar: take the policy's latent state plus raw proprioception and recent sensory history as the mossy-fiber vector x (a few thousand dimensions). Expand through a fixed, sparse, random projection with a threshold nonlinearity into z = \phi(x) with 10^5–10^6 units and ~1–5% activity, with a divisive-normalization "Golgi" term holding sparseness constant. Train only a linear readout y = Wz that adds a correction to the policy's action, at 100 Hz–1 kHz. The teacher is Kawato's feedback-error-learning trick: run a crude feedback controller (or tracking-error signal) in parallel and use its output as the error that updates W — a literal climbing fiber, delayed by ~100 ms as the olive is. The update is a one-line delta rule; no backprop through the foundation model, no replay buffer, adaptation within seconds.

Reuse from Marr: fixed expansion, enforced sparseness, single plastic layer, dedicated error line, capacity accounting. Replace: binary codons with real-valued features, pure random wiring with the redundancy structure connectomics suggests (test whether it buys robustness to dead sensors), and the sign of plasticity per Albus/Ito. The headline measurement: adaptation samples needed to recover performance after a perturbation (payload change, actuator degradation), sidecar vs. LoRA fine-tuning vs. in-context adaptation. My prediction, and I think Marr's: the sidecar wins on sample count and wall-clock by orders of magnitude for the class of errors it can express, and fails gracefully into the slow system for everything else.

Already tried, and what is open

Partially. Albus turned his version into CMAC (1975), which was used in real robot controllers in the 1980s–90s; Kawato's feedback-error learning and cerebellar-inspired adaptive control (e.g., LWPR-era work, Porrill & Dean's adaptive-filter models) are direct descendants; random features and reservoir computing are the ML cousins. What has not been done, as far as I know, is coupling this motif to modern foundation-model policies at scale and measuring the sample-efficiency claim head-to-head — the pieces all exist, the integration does not.

Open on the neuroscience side: what exactly the inferior olive computes (error, timing, or a surprise signal), how much learning lives in the deep cerebellar nuclei rather than cortex, and what the same architecture is doing in the cerebellum's massive projections to prefrontal and language areas — Marr's theory was about movement, but most of the human cerebellum talks to cognition, and nobody has the equivalent theory for that.

Where to read it

The paper: doi.org/10.1113/jphysiol.1969.sp008820 (bibliographic details verified; ~2,700 citations). It is long and idiosyncratic but the codon sections reward slow reading. Read alongside: Albus (1971, Mathematical Biosciences) for the sign correction and CMAC lineage; Ito's LTD work (1982 onward); Litwin-Kumar et al. (2017, Neuron) and Cayco-Gajic & Silver (2019, Neuron) for the modern theory of expansion recoding; Raymond & Medina (2018, Annual Review of Neuroscience) for the honest state of Marr–Albus–Ito; and Kawato's feedback-error-learning papers for the control-theoretic bridge to the sidecar above.