acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Cerebellar Feedback Enables Multi-Task Learning Through Flexible Reorganization of Cortical Dynamics

Abstract

Accumulating evidence suggests that high-dimensional cortical population activity is organized along low-dimensional latent dynamics that support flexible representations and computations. However, existing theoretical models often rely on explicit architectural constraints, such as low-rank recurrent connectivity, to induce structured dynamics, leaving unresolved the question of how such low-dimensional organization may emerge through biologically plausible mechanisms. Inspired by experimental advances in cortico-cerebellar interactions, we propose a cerebellar adaptive feedback mechanism that enables the emergence of low-dimensional cortical dynamics without imposing constraints on recurrent connectivity. We develop a cerebro-cerebellar coordination framework in which a frozen full-rank cortical recurrent neural network (RNN) preserves high-dimensional representational capacity, while an adaptive cerebellar pathway provides task-dependent feedback to regulate cortical dynamics through biologically plausible local plasticity. This interaction reorganizes cortical activity into low-dimensional and separable neural manifolds, reducing representational interference during multi-task learning. We evaluate this framework on cognitive rule learning and handwritten-letter trajectory prediction tasks, demonstrating that adaptive cerebellar feedback effectively improves multi-task learning performance. These findings reveal a biologically grounded computational principle for organizing flexible neural dynamics.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Cortico-cerebellar coordination, Adaptive feedback, Multi-task learning, Dynamical systems, Neural computation

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_QHqnOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a cortico-cerebellar model in which a frozen, full-rank cortical RNN receives adaptive feedback from a cerebellum-inspired module. The authors argue that this feedback reorganizes cortical activity into compact, task-separated representations, thereby reducing interference during multi-task learning.

Critical strengths

The scientific motivation is meaningful. Studying how feedback from another brain region can reorganize cortical dynamics provides a useful perspective on learning. The author perform both single task and multi-task learning. Both models perform well when tasks are trained separately, but their performance differs for multi-task continual learning. Furthermore, The component ablations are useful to show the importance of task-specific feedback.

Critical weaknesses
  1. The authors explain the conceptual difference from Pemberton et al. 2024. A direct comparison would help establish the additional benefits of the proposed plasticity rules and representational reorganization.
  2. Local rules beating gradient descent need explanation. It is surprising that gradient based optimization is about 40 times worse than the local rules.
  3. The training procedure needs clarification. In section C.2, from 994-1004, could the author discuss what parameters are fixed and what are trained or changed using notations you defined?

Minor:

  1. Figure 3A is so difficult to read.
  2. hˉ\bar{h} in equation3 is not defined.
  3. line 798, Figure 7 and Figure S1 are double defined and please check your figure citations.
Questions for authors

Please see weaknesses above.

Reviewer_zY1wOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors train a model of a cortical network coupled to cerebellar feedback learning multiple tasks. Technically, this is cortical RNN with fixed recurrent connectivity, a trained output readout, and adaptive readout and feedback to a cerebellar module. This causes an effective low rank change to the RNN for every learned task. Some of the weights in the network are learned using biologically plausible rules, rather than gradient descent. The authors argue that this setup allows sequential learning of tasks without interference.

Critical strengths

The combination of a cortico-cerebellar loop with plausible learning rules and multi-task training is an important question to study. The apparent lack of catastrophic forgetting is intriguing.

Critical weaknesses

The work of Logiaco, Abbott and Escola (2021) studies a very similar setup and is not mentioned.

The claim that many models engineer low rank structure ignores the literature showing that such structure emerges without engineering it (Schuessler et al 2020; Pellegrino, Cayco-Gajic & Chadwick 2023). Furthermore, the authors use a construction that also constrains connectivity to have a low-rank perturbation via the cerebellar feedback.

There are many details missing, which makes it hard to evaluate the paper (detailed below).

It is not clear how the context-gated activity is implemented. Are the feedback loops from all tasks active at any given time, or does G only select one of them at a time? Unless I missed it, this is a crucial detail that is not properly explained. This is especially important, as the sequential learning of 26 letters is somewhat surprising without a detailed discussion of mechanisms that discourage interference. For instance, is the anti-Hebbian term sufficient to do so? If the gating only enables one task at a time, this could explain much of the effect. But also makes the result less interesting. If not, this requires an explanation of why there is no interference.

Questions for authors

Figure 2: Panel B has no explanation of colors. Panel A bottom has no explanation of different markers. How were the regimes determined?

Line 412: biologically plausible plasticity rules

Lines 412-422: Why do the biologically plausible rules lead to better performance? A naïve expectation is that they would degrade performance compared to a pure optimization approach. This is an interesting observation, but it is neither explained nor explored.

Reviewer_2R2jOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This work proposed a model of multi-task learning in the brain via cortico-cerebellar interaction, where the cortex, modeled as a recurrent neural network (RNN), receives feedback from cerebellum. Using a set of synthetic cognitive and motor tasks, the authors demonstrated that their cortico-cerebellar model achieves sequential multi-task learning, unlike the model without cerebellar inputs. Model ablation indicates that task-specific feedback is crucial for sequential multi-task learning.

The topic addressed in the manuscript, the role of cortico- cerebellar interaction in learning, is an important unresolved question. It is also commendable that the authors implemented their model in two different task settings (synthetic cognitive and motor tasks) and conducted ablations. However, as it is presented, the model appears to have limited contributions. Conceptually, their model doesn’t address the question they posed in the introduction: how to achieve low-rank modification without imposing engineered architectural constraints, because cerebellar feedback is assumed to be low rank. This assumption is not well supported from the biological viewpoint either, as the number of granule cells in the cerebellum is larger than the population size of cortical regions. The task signal-based gating they introduced to resolve sequential multi-task learning is also an established method in the continual learning literature.

Critical strengths

None

Critical weaknesses

Low-rank feedback: While the introduction section states that the goal is to achieve a flexible and adaptive realization of low-dimensional control, the proposed model explicitly assumes low-rank feedback, falling short of addressing this goal. It also makes the novelty over the previous models that similarly assumed low-rank plasticity unclear. Additionally, the model assumption of low-rank cerebellar feedback is not well motivated biologically, because the cerebellar granule cell layer contains a large number of neurons, and thus, the feedback to cortex is unlikely to be low rank.

CF gating: In sequential multi-task learning, the authors assumed that climbing fibers (CF) provide contextual gating signals based on the task identity. Firstly, this is a well-established mechanism (e.g., Serra et al., PMLR 2018; Masse et al., PNAS 2018), and thus it is not surprising to observe successful sequential learning in this setting. Secondly, while CF carries some contextual signals, its implementation seems rather unnatural. The model assumes that CF input purely conveys context signals, and the error is provided in parallel from an unknown source. This is at odds with the established role of CF input in error-driven learning.

Model assumptions: The presented modeling appears ad hoc. The local plasticity rules were introduced without clear justification, and the variations of those rules were not explored beyond comparison with the Adam optimizer. While the paper motivates its modeling choice from biological plausibility, the plausibility should be questioned in several aspects (see comments above). The manuscript offers a phase diagram analysis, but it follows rather trivially from a previous work (Mastrogiuseppe and Ostojic, 2018).

Serra, Joan, et al. "Overcoming catastrophic forgetting with hard attention to the task." International conference on machine learning. PMLR, 2018.

Masse, Nicolas Y., Gregory D. Grant, and David J. Freedman. "Alleviating catastrophic forgetting using context-dependent gating and synaptic stabilization." Proceedings of the National Academy of Sciences 115.44 (2018): E10467-E10475.

Mastrogiuseppe, Francesca, and Srdjan Ostojic. "Linking connectivity, dynamics, and computations in low-rank recurrent neural networks." Neuron 99.3 (2018): 609-623.

Questions for authors

[Clarification] I was surprised that, in Table 2, all-gradient didn’t work. Can you expand on the observed failure mode? Was the convergence too slow, or did it suffer from catastrophic forgetting?

[Suggestion] Have you tried imposing sparsity directly onto c(t)c(t)? Seem to me, this is more biologically sensible way to introduce sparsity beneficial for sequential multi-task learning.

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

The paper proposes a brain inspired model in which a cerebellum module helps a cortex network learn several tasks one after another without forgetting earlier ones. The cortex is a recurrent network whose internal weights are frozen, only its output layer is trained. The cerebellum module reads the cortex’s activity and sends back a feedback signal. Each task gets its own low-rank feedback (U_k, V_k, g_k), learned with brain-like local learning rules instead of backprop.

On four cognitive tasks learned in sequence (DelayPro -> DelayAnti -> MemoryPro -> MemoryAnti), the cortex-only model forgets earlier tasks, while the coupled model keeps all four. The coupled model also produces lower-dimensional, better-separated neural activity (Figs. 3C, 4C). On a harder benchmark of 26 handwritten letters learned in sequence, the coupled model retains the alphabet while the cortex only model forgets quickly (Fig. 6). Ablations test which parts of the feedback and which learning rules matter (Tables 1-3).

Critical strengths
  1. Using a cerebellum like feedback loop, trained with local plasticity rules rather than backpropagation, to shape cortical dynamics is a well motivated idea. It works on two different benchmarks, including a 26-letter sequential task.
  2. Tables 1-3 test which parts of the feedback matter (shared vs. task-specific parts), which learning rules matter, and whether a fixed low-rank structure alone is enough. This helps isolate what drives the results.
Critical weaknesses
  1. The coupled model gets its own feedback parameters for each task (U_k, V_k, g_k), chosen using the known task identity. The cortex only model has only one shared output layer that is retrained on each new task. Giving each task its own parameters is a known way to prevent forgetting, so the large gap in Table 3 (0.0029 vs. 0.3160) is expected.
  2. The cortex only baseline trains only its output layer, which makes it very weak. Even fully shared feedback (Table 1) improves MSE from 0.316 to 0.0195, suggesting much of the gain comes simply from adding trainable feedback.
  3. There is no comparison with standard continual learning methods.
  4. In Table 2, replacing the local rules with gradient training (Adam) makes performance much worse (0.1231 vs. 0.0029). Gradient training would normally do at least as well. Is this because the gradient version is poorly tuned.
Questions for authors
  1. How does the model compare with a cortex only network that also gets task specific parameters, for example a separate output layer etc?
  2. How does it compare with standard continual learning methods?
  3. Why does gradient training perform so much worse than the local rules in Table 2? How were its settings tuned?
  4. How many seeds were used?
  5. How does the model know which task’s feedback (U_k, V_k, g_k) to use? Is it biologically plausible to keep separate feedback parameters for every task?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.