acceptodds
Under review as a conference paper at ICLR 2027

A Plasticity Model of Cerebellar Regulation of Cortical Dynamical Stability and Task-Associated Representations

Abstract

Experiments indicate that cerebellar regulation helps stabilize cortical activity, but the mechanisms that support this effect remain unclear. We construct a biologically motivated cortical–cerebellar model to study how cerebellar learning can regulate an active cortical circuit. A recurrent neural network (RNN) with fixed connections and active/chaotic dynamics represents the cortical circuit, while continuous online learning acquires cerebellar feedback. Differential Hebbian plasticity at the input stage associates cortical pattern activity with category-related cues (context) supplied by an auxiliary cortical population; the output stage uses eligibility traces and a teaching signal that reflects the history of internal activity changes to learn feedback. In a five-category pattern task, removing context after training leaves cerebellar feedback able to reduce trajectory speed and repeated-stimulus perturbation distance while increasing category separation at the final state. Structural replacement and plasticity ablations show that the expanded sparse feedforward cerebellar pathway and coordinated adaptation at the input and output sites jointly support response stabilization and category organization. When category cues are mismatched, part of the dynamical control is preserved but category separation is impaired; unlearned categories evoke stronger responses and weaker cumulative perturbation contraction. These results suggest that the cerebellum can learn to regulate effective recurrent drive by sensing and integrating internal cortical changes, while associative plasticity allows stabilization to retain task-related pattern structure.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Recurrent Neural Networks, Chaotic Dynamics, Differential Hebbian Learning, Eligibility Traces, Associative Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_21mHOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a network inspired by cortico-cerebellar loop. The "cerebellum" network learns to regulate a fixed recurrent "cortical" network using local plasticity rules. Training combines label cues with a teaching signal computed from activity history. On a five-class MNIST benchmark, the authors report reduced activity drift, greater consistency across repeated responses, and improved class separation after the cues are withdrawn. The results show architecture feasibility within the chosen setup and provide a computational model to explain how cerebellar learning regulates cortical dynamics.

Critical strengths

The paper offers an interesting perspective on cerebellar feedback as a mechanism for shaping cortical dynamics and task representations.

The question of how coupled brain systems maintain stable activity while preserving task information is important and worth investigating. The biological motivation makes this a compelling direction for computational modeling.

Critical weaknesses
  1. Poor naming choices... The writing is poor, and the logic is not well organized. I struggled to grasp the main idea. Many of the terms are confusing, unintuitive, and sometimes even misleading.

  2. Unmatched model capacity. A 1,000-unit recurrent core and 32,000-unit expansion, with 32 million output connections, are insufficiently justified for this five-category benchmark. Missing size sweeps and capacity controls leave Table 1 confounded by model size and simultaneous architectural changes.

  3. Unestablished convergence. Figure S1 is still improving at the final checkpoint, and corresponding curves are missing for the comparison conditions. Tables 1 and 2 therefore do not distinguish learning speed from performance after sufficient training.

  4. Insufficient mechanistic support. The hand-designed updates lack sufficient analysis explaining their joint effects on stability and class organization. And there's no explicit optimization objective supporting the learning rule.

Questions for authors
  1. What justifies the 1,000-unit cortical network and 32,000-unit expansion? Do the findings hold with smaller networks or capacity-controlled alternatives?

2.Do Tables 1 and 2 characterize fixed-budget learning or converged performance? Could you provide some more figures showing the training process?

3.What objective or local control principle motivates the two learning rules, and why should they jointly improve stability while preserving class information?

Reviewer_8RAPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates how cerebellar plasticity can stabilize cortical dynamics and generate context-dependent representations. This is a timely and important topic, particularly given recent experimental findings highlighting the role of granule cells in context-dependent associations.

Critical strengths
  1. The model provides a computational framework for investigating the interaction between associative plasticity in the granule cell layer and cortical dynamical stability.

  2. The results show improvements in cortical stability and category separation within the evaluated setting, providing an initial demonstration of the proposed mechanism.

Critical weaknesses
  1. Biological plausibility. The authors emphasize learning without supervised target outputs or trajectory errors as a major advantage. However, previous cortico-cerebellar models (e.g., Pemberton et al., Agueci and Cayco-Gajic) use sensory or performance-related teaching signals supported by experimental evidence. Purkinje cells encode movement-related variables, and climbing fibers from the inferior olive convey instructive signals, including sensory prediction errors. Although cortical activity could contribute to these signals, replacing them with an internally computed cortical signal does not necessarily improve biological plausibility. Can you better justify this?

  2. Contextual inputs. The auxiliary cortical population providing category-dependent contextual signals is insufficiently motivated. The biological origin of these signals, the choice of preferred angles, and their contribution to learning should be clarified through appropriate ablations.

  3. Novelty and model complexity. Previous biologically motivated models, particularly Pemberton et al. (2024) and Agueci and Cayco-Gajic (bioRxiv), have already addressed task acquisition, switching, and consolidation using simpler architectures. The authors should clarify their contribution relative to these approaches and demonstrate why the additional components are necessary, ideally through direct comparisons and targeted ablations.

  4. Limited experimental validation. The model is evaluated primarily in a single MNIST setting. Although the results demonstrate stabilization and category separation, the generality of the mechanism remains unclear. Testing sequential acquisition of new categories without disrupting previously learned representations would help establish its relevance to flexible and continual learning.

  5. Finally, the manuscript is a bit dense and difficult to follow, and the figures could be improved to help clarify the proposed principles.

Questions for authors

See above.

Reviewer_GyMnOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces a cortico-cerebellar model consisting of a fixed cortical recurrent network and a plastic cerebellar feedforward pathway characterised by high-dimensional expansion and sparse coding. The model is evaluated on a simple classification task using the first 5 digits from MNIST, using 400 training samples per class as training, and with each trial lasting 1 second. Performance is evaluated with three primary metrics: the mean silhouette in the original cortical activity space (measuring category organisation), trajectory speed (measuring rate of change in activity), and repeated-stimulus perturbation distance D(1000) (measuring the separation after 1s between responses to the sample sample with two independent iperturbed initial states).

The authors analyse the stability of cortical responses and preservation of input category structure after associative input is withdrawn, comparing the complete model against a cortex-only version of it. The complete model exhibits improved stability and separation with respect to the ablated version. The cerebellar feedforward pathway is also compared with a cortex-like recurrent replacement module (with dense input, continuous hidden activity, and recurrence) using the same cortical substrate and training protocol. The feedforward protocol was demonstrated to achieve higher category separation and lower perturbation distance. The authors subsequently show that category organisation in their network is relatively robust to small cue noise, but is disrupted by substantial mismatch. Finally, they examine responses to out-of-distribution classes, showing that they evoke stronger internal-change responses and weaker mean cumulative contraction in the network.

Critical strengths

The authors introduce a biologically-inspired cortico-cerebellar model that demonstrates several potentially useful properties, especially with respect to the stabilisation of cortical dynamics and preservation of category-related representations. Although an initial characterisation of such properties is provided by the experiments on MNIST,the evaluation is restricted to a single and relatively simple task.

Critical weaknesses

While the Introduction provides a neurocientific motivation for the proposed architecture, the connection between the results obtained by the model and experimental neuroscience is very narrow throughout the text. The Discussion provides a brief and under-developed interpretation relating results back to neuroscience (line 479).

In the main text, the authors introduce and present multiple quantitative measures. However, the explanations of such quantities are often missing or very concise and, in my opinion, difficult to interpret; sometimes even cryptic. Some representative examples are:

  • "[...] we compared paired pattern stimulation with the cortex-only condition and with cerebellar feedback retained. Late-period trajectory speed decreased from (10.163±1.044)×10−3 ms−1 in the cortex-only condition to (0.434±0.237)×10−3 ms−1 with feedback, while the repeated-stimulus perturbation distance D(1000) decreased from 0.750 ± 0.062 to 0.024 ± 0.014 (Figure 2a,b)." Figure 2c is then presented, missing an interpretation.
  • "To distinguish the contributions of the two sites during training, we froze P or W [...]. The final silhouette coefficients after freezing P were 0.128±0.032 and 0.191±0.071, respectively; after freezing W they were 0.287± 0.094 and 0.437 ± 0.084 (Figure 3a)." Figure 3b is then presented, missing an interpretation.
  • "[...] we compared learned (in-distribution, ID) classes 0–4 with unlearned (out-of-distribution, OOD) classes 5–9. The ID and OOD final-time responses were 0.313 ± 0.020 and 0.428 ± 0.014, respectively, and the area under the receiver operating characteristic curve (AUROC) was 0.706 ± 0.035 (Figure 4a)." Figure 4b is then presented, missing interpretation.

These results could be informative to a machine, but they are barely discussed, interpreted, wired back to neuroscience throughout the paper. The "Discussion" Section reiterates the main findings rather than giving a critical examination of their implications, or providing a broader interpretation that the reader would expect in the main text.

The "Discussion" Section includes only a very brief mention of the network's limitations and possible future directions, three lines in total (line 483). These should be particularly relevant given the simple nature of the proposed experimental setup.

The authors do not commit to releasing their code.

Questions for authors

I do not have any specific question for the authors, see weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.