{"id":"ebafb674-08c0-43f8-819d-fafd63b81833","arxiv_id":"2505.14355","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":0.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A review of how perceptron theory has shaped cerebellar learning research, from Marr and Albus to modern plasticity experiments.","lead":"This paper reviews how the perceptron, a simple neural network model, has been used to understand how Purkinje cells in the cerebellum learn movements. It covers the classic Marr-Albus-Ito theory, recent experimental tests, and open questions about credit assignment and temporal coding.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the paper's central claim is a descriptive account of Marr-Albus-Ito theory, and its acknowledged limitations do not undermine that account.","rationale":"The reader's verdict is ACCEPT, and I agree. The paper is a review article, so the relevant standard is accuracy and fair representation of the literature, not novel proof. The reader's strongest_claim interprets the paper as asserting that the cerebellum is a supervised learning machine, but the paper's own wording is more careful: 'Marr-Albus-Ito theory now usually refers to the view...' That is a descriptive claim about the field's terminology, not a proof that the theory is complete. The weaknesses the reader identifies as load-bearing are explicitly acknowledged in the paper: the independence assumption is called 'unrealistic,' and the climbing-fiber-as-error-signal hypothesis is described as difficult to answer, with evidence of reward and behavioral signals cited. The review also states that there is no evidence for gradient-descent-like learning in the brain and that going 'beyond Marr-Albus-Ito theory' is a future goal. These admissions are consistent with a balanced review and do not create an internal inconsistency. I find no significant objection to the central claim as actually stated; the only check worth running is a faithfulness check of the historical attribution, since the review's value depends on accurate reporting of Marr and Albus's original proposals. The verdict should remain UNCHANGED.","tokens_in":15099,"tokens_out":3705,"duration_ms":36664,"concrete_test":"Check the faithfulness of the historical claims by comparing the paper's description of Marr 1969 and Albus 1971 with the original texts, in particular the n-uple recoding argument and the claim that Marr proposed PF-PC potentiation; if either description is inaccurate, the review's historical framing would need correction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"No load-bearing concern found. The paper is a review and its central assertion is explicitly about what Marr-Albus-Ito theory 'usually refers to' rather than an unqualified claim that the cerebellum is a supervised learning machine. The independence assumption in perceptron capacity calculations is flagged by the authors themselves, and the review likewise reports that climbing fibers carry reward and behavioral signals and that perturbation-based credit assignment 'remains to be demonstrated in vivo.' These are open questions, not internal inconsistencies. The review's own concluding section forecasts going 'beyond Marr-Albus-Ito theory,' which is compatible with describing it as the organizing framework. I therefore see no argumentative step that would need to be fixed for the paper to fulfill its stated purpose.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This review article, written in honor of Gérard Toulouse, traces the intellectual path from the perceptron to the cerebellum. It recounts how Marr and Albus cast Purkinje cells as perceptrons supervised by climbing-fiber teaching signals, with granule cells providing a high-dimensional expansion of mossy-fiber inputs. The paper reviews the statistical-physics and Gardner-style capacity calculations for perceptrons with sign-constrained weights, the comparison of the predicted synaptic weight distribution to experimental Purkinje-cell data, the extension to temporal basis functions for delayed and timed responses, and the modern debate about whether climbing fibers carry a dedicated per-cell error signal or also reward and behavioral information. It closes by identifying open questions, including credit assignment in complex movements, the functional role of cerebellar rhythms, and the extent to which cerebellar learning goes beyond the classical Marr-Albus-Ito framework.","tokens_in":15345,"tokens_out":2534,"duration_ms":26434,"significance":"As a review for a special issue commemorating Gérard Toulouse, the paper serves a clear and appropriate purpose: it explains how the perceptron, a central object in the statistical physics of neural networks, became a foundational model for cerebellar motor learning and how this perspective remains influential. Its significance lies in synthesis rather than new results, but the synthesis is accurate, balanced, and current. The authors explicitly acknowledge the main limitations of the classical theory, including the unrealistic statistical-independence assumption in capacity calculations and the evidence that climbing fibers are not purely error signals. The review is particularly valuable for readers outside the cerebellum field who want a condensed, authoritative account of the Marr-Albus-Ito theory and its modern variants, including the perturbation-based credit-assignment proposal of Bouvier et al. It also gives appropriate credit to the historical role of Toulouse and his collaborators in importing statistical physics ideas into neuroscience.","major_comments":[],"minor_comments":[{"comment":"The caption of Fig. 1A labels 'molecular layer interneurons (UBC)', which conflicts with the standard abbreviation for unipolar brush cells and with the text that treats molecular layer interneurons and UBCs as distinct cell types; the caption should be corrected to 'molecular layer interneurons (MLI)'.","section":"Fig. 1 caption"},{"comment":"The sentence 'Numerous experiments, notably from Thomson's lab [22]' contains a typo: the laboratory is that of Richard Thompson, so 'Thomson's lab' should be 'Thompson's lab'.","section":"Section II"},{"comment":"The phrase 'the presence of absence of a second complex spike' should read 'the presence or absence of a second complex spike'.","section":"Section VI"},{"comment":"Several typographical errors should be corrected: 'representions' in Section III, 'paralell' in Section IV, 'Purkjinje' in Section VI, and 'seeked' in Section I.","section":"Throughout"},{"comment":"The name 'Cayco Gajic' is written inconsistently (without a hyphen in the text, with a hyphen in the reference list); it should be unified to 'Cayco-Gajic'.","section":"Section III and reference list"},{"comment":"The phrase 'Electrophysiological experiments inin vitro preparations' contains a doubled 'in'; it should read 'in in vitro preparations' or 'in vitro preparations'.","section":"Section II"},{"comment":"Several references contain character-encoding artifacts (e.g., 'p. 960ˆ a€“962', '38–55' garbled as 'p. 38ˆ a€“55'): these should be cleaned in the final typeset version.","section":"Reference list"}],"recommendation":"minor_revision","confidential_remarks":"This is a well-written review appropriate for the commemorative issue. The central claims are descriptive and appropriately hedged, and the acknowledged limitations do not undermine the paper's purpose. The main work before publication is copyediting: fixing the figure caption error, the 'Thompson' typo, and several minor typographical issues. The heavy citation of the authors' own work is understandable in a historical review that covers their own contributions, and it does not distort the account."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a review, not a research paper, and it is a good review. It traces the perceptron-to-cerebellum arc from Marr and Albus through Gardner's capacity calculations to modern work on expansion recoding, temporal basis functions, and credit assignment. If you want one readable account of where Marr-Albus-Ito theory stands and where it wobbles, this is it.\n\nWhat the paper does well: it is accurate on the biology and the theory. The geometric argument for granule-cell expansion is clearly explained, the perceptron capacity results are correctly summarized, and the comparison of predicted silent-synapse fractions with Isope and Barbour's data is presented fairly. It also gives real credit to alternative views: climbing fibers carry reward and behavioral signals, not just error; the perturbation-based credit-assignment scheme is explicitly flagged as awaiting in vivo demonstration; and the independence assumption in standard capacity calculations is acknowledged as unrealistic, with correlated-input extensions cited. The conclusion is appropriately modest about going beyond Marr-Albus-Ito, which is the right note for a review that is also a tribute.\n\nThe soft spots are minor. Figure 1's legend says \"molecular layer interneurons (UBC)\" where it should say \"MLI\" or \"basket/stellate cells\"—an obvious typo, but the kind that confuses readers. \"Thomson's lab\" is a typo for Thompson's lab. The paper is also visibly authored by people who did much of the work being reviewed, so self-citation is frequent; here that is appropriate rather than inflationary, since the cited papers are the ones that actually established the results.\n\nThe only real concern, if you want one, is that the paper does not critically discriminate between the strong claim that the cerebellum is a supervised learning machine and the weaker claim that Marr-Albus-Ito theory is the best current framework. But the authors themselves use the weaker framing throughout, so this is not a flaw in execution.\n\nWho is this for? A reader who wants a compact, reliable entry point into the theory of cerebellar learning, or a physicist curious how perceptron ideas migrated into neuroscience. It is not a research contribution and makes no pretense of being one. I would send it to peer review and accept after the typos are fixed.","headline":"A solid, honest review of Marr-Albus-Ito theory that consolidates the field and flags its own soft spots; nothing new, but it deserves a serious referee and probably acceptance after minor fixes.","tokens_in":15735,"tokens_out":1284,"would_cite":true,"duration_ms":13598,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that Marr-Albus-Ito theory remains the organizing framework for cerebellar motor learning.","keywords":["cerebellum","perceptron","Marr-Albus-Ito theory","Purkinje cell","granule cells","climbing fiber","synaptic plasticity","supervised learning"],"falsifier":"A falsifying observation would be a motor-learning experiment in which complex spikes are recorded during a well-controlled task and shown to encode reward expectation or movement rather than error, or in which blocking complex spikes does not prevent learning; either would break the teacher role. Alternatively, large-scale imaging showing dense, low-dimensional granule-cell activity during natural behavior would undercut the expansion argument.","tokens_in":14907,"feed_emoji":"🧠","tokens_out":9497,"duration_ms":85225,"temperature":0.7,"pith_summary":"This review argues that the Marr-Albus-Ito theory remains the organizing framework for cerebellar motor learning. The core idea is that a Purkinje cell acts as a perceptron: granule cells recode mossy-fiber inputs in a much higher-dimensional, sparser space, and a climbing fiber from the inferior olive supplies the teacher that depresses active parallel-fiber synapses. The paper traces this proposal from its origin through statistical-physics calculations of perceptron capacity, showing how the theory predicts the measured distribution of silent versus active synapses and how it extends to correlated inputs, temporal basis functions, and a perturbation-based solution to credit assignment. A sympathetic reader comes away with the cerebellum as a concrete biological example of a supervised learning machine.","feed_headline":"Cerebellar learning still fits the perceptron model","feed_subtitle":"Purkinje cells act as perceptrons, granule cells expand inputs, and climbing fibers carry the teacher signal.","key_machinery":"The central object is the sign-constrained perceptron model of a Purkinje cell, with $N$ parallel-fiber inputs, non-negative synaptic weights, and a threshold on the summed input; the learning rule is depression at active parallel-fiber synapses when the climbing fiber fires. Two quantities carry the argument: the expansion ratio of the granule-cell representation, which makes overlapping mossy-fiber patterns nearly orthogonal and linearly separable, and the critical capacity $\\alpha_c = p_{\\max}/N$ from statistical-physics analyses, which predicts that at maximal capacity a large fraction of synapses should be silent, matching the distribution seen in paired recordings.","core_discovery":"The paper's central claim is that the cerebellar cortex should be understood as a supervised learning machine of the perceptron type: Purkinje cells are the output units, granule cells provide an expanded representation that makes input patterns easier to separate, and each climbing fiber acts as a teacher that reports error and depresses active parallel-fiber synapses. It consolidates the evidence by combining anatomical numbers (about 100,000 parallel-fiber inputs per Purkinje cell), the combinatorial recoding argument for expansion, classical linear-separability results, statistical-physics capacity calculations for sign-constrained weights, and quantitative fits to paired electrophysiological recordings. The review also presents later extensions: learning correlated input-output sequences, analog perceptrons, temporal basis functions for properly timed responses, and a perturbation-based algorithm in which climbing fibers serve both as exploratory noise and as an error feedback signal. The contribution is not a new experiment but a defense of Marr-Albus-Ito theory as the organizing account of cerebellar motor learning.","pith_inferences":["If climbing fibers turn out to carry reward-related signals as often as error signals, the perceptron metaphor would need to be broadened to a reinforcement-learning rule, preserving Marr-Albus-Ito as a family of supervised and perturbation learners rather than a strict error-teacher account.","The expansion principle may be a general circuit motif: the same expansion-and-sparsification design appears in hippocampal and insect olfactory relays, so the optimality result for four to five inputs per expansion cell suggests a testable common design rule across systems.","The independence limitation points to a natural experiment: record natural mossy-fiber activity during a learned behavior and measure how much capacity a Purkinje cell retains when input correlations match those of the behavior.","The perturbation-based algorithm implies that spontaneous complex spikes before learning act as exploratory motor perturbations, a prediction that could be tested by tracking trial-to-trial movement variability against complex-spike timing."],"forward_implications":["The high-dimensional, sparse recoding by granule cells should make mossy-fiber patterns linearly separable, giving a functional reason for the large expansion between mossy fibers and parallel fibers.","At maximal capacity with sign-constrained weights, the perceptron predicts a large fraction of silent parallel-fiber synapses, matching the roughly 80 percent silent contacts reported in paired recordings.","Learning statistically correlated input-output sequences can raise capacity, and bistability helps further when output correlations exceed input correlations.","Temporal basis functions from unipolar brush cells, diverse mossy-fiber time scales, and short-term plasticity allow Purkinje cells to produce correctly timed outputs even after the sensory input has ended.","A perturbation-based algorithm in which climbing fibers both perturb movements and signal success reproduces observed plasticity rules and offers a candidate solution to the cerebellar credit-assignment problem."],"supporting_citations":[{"why":"Proposes that granule cells recode mossy-fiber patterns in a higher-dimensional space and treats Purkinje cells as perceptrons.","marker":"[18]"},{"why":"Supplies the alternative proposal that the climbing fiber is a teacher and that joint parallel-fiber and climbing-fiber activity depresses active synapses.","marker":"[19]"},{"why":"Provides the experimental evidence that climbing-fiber stimulation depresses parallel-fiber responsiveness.","marker":"[20]"},{"why":"Establishes that the number of linearly separable patterns grows linearly with input dimension, justifying the expansion benefit.","marker":"[27]"},{"why":"Provides the statistical-physics method for computing the critical capacity of a perceptron with random patterns and sign-constrained weights.","marker":"[47, 48]"},{"why":"Applies the sign-constrained perceptron to Purkinje cells and fits the predicted silent-synapse distribution to experimental recordings.","marker":"[42]"},{"why":"Shows that four to five mossy-fiber inputs per granule cell nearly maximize the dimensionality of the granule-cell representation.","marker":"[35]"},{"why":"Extends the perceptron model to correlated input-output sequences and bistable Purkinje cells, changing the predicted capacity.","marker":"[51]"},{"why":"Proposes the perturbation-based algorithm in which climbing fibers act as both noise and error feedback, matching in vitro plasticity outcomes.","marker":"[68]"}],"fun_headline_variants":["Perceptron model still explains cerebellar learning","Cerebellum as a perceptron: Marr-Albus theory lives","Purkinje cells as perceptrons: 100k inputs per cell","Climbing fibers: teachers in the cerebellar perceptron"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each climbing fiber delivers a dedicated per-cell error signal and that the associations a Purkinje cell must learn are statistically independent; if real climbing fibers mainly carry reward or behavioral information, or if mossy-fiber inputs are strongly correlated, the capacity and credit-assignment arguments shift.","fun_headline_variants_meta":{"raw":{"variants":["Perceptron model still explains cerebellar learning","Cerebellum as a perceptron: Marr-Albus theory lives","Purkinje cells as perceptrons: 100k inputs per cell","Climbing fibers: teachers in the cerebellar perceptron"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000375,"raw_usage":{"total_tokens":1930,"prompt_tokens":804,"completion_tokens":1126,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":420,"completion_tokens_details":{"reasoning_tokens":1056}},"tokens_in":420,"tokens_out":1126,"duration_ms":10535,"temperature":1.0,"reasoning_tokens":1056,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:34:24.667836+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A falsifying observation would be a motor-learning experiment in which complex spikes are recorded during a well-controlled task and shown to encode reward expectation or movement rather than error, or in which blocking complex spikes does not prevent learning; either would break the teacher role. Alternatively, large-scale imaging showing dense, low-dimensional granule-cell activity during natural behavior would undercut the expansion argument.","supporting_citations":[{"cited_title":"A theory of cerebellar cortex,","cited_arxiv_id":null,"evidence_quote":"Proposes that granule cells recode mossy-fiber patterns in a higher-dimensional space and treats Purkinje cells as perceptrons."},{"cited_title":"A theory of cerebellar function,","cited_arxiv_id":null,"evidence_quote":"Supplies the alternative proposal that the climbing fiber is a teacher and that joint parallel-fiber and climbing-fiber activity depresses active synapses."},{"cited_title":"Climbing fibre induced depression of both mossy fibre responsiveness and glutamate sensitivity of cerebellar Purkinje cells,","cited_arxiv_id":null,"evidence_quote":"Provides the experimental evidence that climbing-fiber stimulation depresses parallel-fiber responsiveness."},{"cited_title":"Geometrical and statistical properties of systems of linear inequalities with applications in pattern recogni- tion.,","cited_arxiv_id":null,"evidence_quote":"Establishes that the number of linearly separable patterns grows linearly with input dimension, justifying the expansion benefit."},{"cited_title":"Optimal information storage and the distribution of synaptic weights: perceptron versus Purkinje cell.,","cited_arxiv_id":null,"evidence_quote":"Applies the sign-constrained perceptron to Purkinje cells and fits the predicted silent-synapse distribution to experimental recordings."},{"cited_title":"Optimal degrees of synaptic connectivity,","cited_arxiv_id":null,"evidence_quote":"Shows that four to five mossy-fiber inputs per granule cell nearly maximize the dimensionality of the granule-cell representation."},{"cited_title":"Storage of correlated patterns in standard and bistable Purkinje cell models,","cited_arxiv_id":null,"evidence_quote":"Extends the perceptron model to correlated input-output sequences and bistable Purkinje cells, changing the predicted capacity."},{"cited_title":"Cerebellar learning using perturbations,","cited_arxiv_id":null,"evidence_quote":"Proposes the perturbation-based algorithm in which climbing fibers act as both noise and error feedback, matching in vitro plasticity outcomes."}],"review_version":1}