acceptodds
Under review as a conference paper at ICLR 2027

Efficiency and robustness partition the solution space for memory in recurrent networks

Abstract

In computational neuroscience, task-trained recurrent neural networks (RNNs) are commonly used as a testbed for exploring the space of recurrent circuit solutions compatible with a particular function. However, these networks are subject to inductive biases that may be misaligned with those of biological circuits, which operate under various efficiency and robustness constraints. Here, using a minimal stimulus recall task, we systematically characterize how the solution space of recurrent neural networks is shaped by the desiderata of weight efficiency, activity efficiency, and robustness to noise. We show that weight efficient solutions exhibit a low-rank bias that aligns with the inductive biases of task-trained networks and favors dynamical motifs involving slow, persistent dynamics, whereas activity efficiency encourages high-rank solutions that rely on transient amplification. Moreover, we find that activity efficient solutions exhibit a strong dissociation between high-variance modes of activity and those that causally influence network output. Unlike in static feedforward settings, we demonstrate that noise robustness and activity efficiency are generally in opposition, rather than synergistic. Using these findings, we propose a normative account of misaligned coding phenomena recently identified in premotor cortex.

Primary area:
applications to neuroscience & cognitive science
Keywords:
recurrent neural networks, computational neuroscience, dynamical systems

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_LdsROpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors study analytically the space of solutions of a simple learning problem in a linear rnn (plus a nonlinear case numerically) from the perspective of activity and weight regularization. They characterize two extreme cases where weight or activity cost dominate, and the resulting structure is rank-1 or high-rank, effectively feed-forward, respectively. They validate their findings numerically, compare with the implicit regularization induced by gradient learning, and use an interesting measure of noise-sensitivity based on adjoint states.

Critical strengths
  • The setting is controlled and the problem formulation is very clear

  • The authors use the lambda->0 and lambda>>1 extreme cases to nicely interpolate the other optimal solutions; the emergence of an effective-feed forward structure is explained clearly, as is the conservation law explaining the efficiency/noise-robustness trade-off

Critical weaknesses
  • The novelty wrt Ritter & Chadwick and similar works about low-rank vs feed-forward solutions in rnns needs to be discussed more clearly.

  • the link with signal-to-noise ratio optimization is briefly discussed before equation (7) but should be expanded a bit more

Questions for authors
  • how would you expect the results to change with multiplicative noise?

  • it would be nice to have an additional characterization of the misaligned-coding phenomenon across lambda and gamma, it looks like the effect is way less evident even for a very small adjoint cost

2 comments

We thank the reviewer for their helpful comments and questions, as well as for their appreciation of the merits of our work. We address the points raised under Weaknesses and Questions below:

Novelty wrt Ritter & Chadwick:

We have now expanded the discussion section with an additional sentence clarifying how our paper's findings relate to, yet are substantially distinct from, those of Ritter & Chadwick in lines 482-485: "While [Ritter and Chadwick] also explore the effects of activity regularization numerically, their analytical treatment focuses primarily on characterizing the signal-to-noise performance of working memory networks with different recurrent connectivity patterns, rather than how competing biological costs might preferentially select among them."

SNR optimization instead of absolute noise scale

We have now added a sentence (lines 300-303: "Indeed...variance.") clarifying that for the problem we study, SNR maximization is equivalent to minimizing the absolute noise sensitivity (adjoint cost) that we introduce. Since this point also overlaps with one raised by reviewer 2dwf, we also direct the reviewer's attention to the following point we raised there, repeated below:

There is a fairly direct way to massage the tradeoff inequality in Eq. (6) into one relating SNR to activity cost. If the gain g(τ)g(\tau) is allowed to be optimized as well (otherwise SNR is effectively identical to inverse absolute noise), then we straightforwardly obtain by similar arguments that ExEq(τ)≥g(τ)2τ2E_x E_q^{(\tau)} \geq g(\tau)^2\tau^2, where Eq(τ)=∫0τ∥q(t)∥2dtE_q^{(\tau)} = \int_0^\tau \|\mathbf{q}(t)\|^2 dt, and thus can upper bound SNR(τ)=g(τ)2ση2Eq(τ)\text{SNR}(\tau) = \frac{g(\tau)^2}{\sigma^2_\eta E_q^{(\tau)}} as SNR(τ)≤1ση2τ2Ex\text{SNR}(\tau) \leq \frac{1}{\sigma^2_\eta \tau^2}E_x. Morally, this is expressing a similar fundamental constraint as what we state: to achieve high SNR, one needs to pay in activity cost.

Multiplicative noise:

Under the most natural/simple model of multiplicative noise, where the input noise entering the model is given by x(t)⊙η(t)\mathbf{x}(t) \odot \boldsymbol{\eta}(t), we suspect that this would weaken, but not entirely ablate, the tradeoff between activity efficiency and noise robustness. This is because lower activity would also reduce the scale of the injected noise, thereby making noise robustness and activity efficiency more ``aligned" as objectives. We leave this question to future work.

Effect of (λ\lambda, γ\gamma) on misaligned coding:

The reviewer is correct in that introducing an adjoint cost can substantially dampen the misaligned coding phenomena observed, as reported in Fig. E.1. However, based on analysis of both the idealized linear setting and in the nonlinear ALM experiments, the emergence of at least some degree of misaligned coding with small λ\lambda remains robust to the introduction of process noise. This remains true whether the noise is introduced indirectly through the γ\gamma-weighted adjoint cost or directly through adding noise to the dynamics itself, as was done in the ALM experiments. Nonetheless, we agree that a more detailed investigation of the entire (λ,γ)(\lambda , \gamma) plane and its modulation of misaligned coding phenomena would also be a useful direction for future work.

I'm happy to recommend acceptance.

Reviewer_C3pEOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies how different requirements (weight efficiency, activity efficiency, and noise robustness) shape the solution space of RNNs. A linear RNN is trained on a simple stimulus recall task under constraints on activity and weight magnitude. The authors analyze the impact of these constraints on network dynamics, find that different constraints lead to distinct solutions and compare these with biological networks solving similar tasks. The paper is addressing an interesting question with analytical tools, is very well-written and I think it should be accepted for publication in the conference.

Critical strengths

S.i) Interesting question: The topic is very interesting. The question of how different regularizers can drive the RNN to find different solutions is important both for people working on artificial systems as well as neuroscience research using RNNs as a model for brain dynamics. The analysis of the low-rank bias of RNNs and link to the weight efficiency as well as the contrast between the feed forward and recurrent results are genuinely useful results.

S.ii) Real interpretable analytics. The setup is simple enough to produce real analytical results. The choice of regularizers is intuitive and by just limiting analysis to a few parameters (λ, γ) the results become explainable. Th. Additionally, the test on the nonlinear RNN and the ALM are thorough and support the conclusions well.

Critical weaknesses

I see no major weaknesses. The main limitations are acknowledged by the authors and are raised as questions below. In general I think its a really nice paper that can be accepted with minor edits.

Questions for authors

This is a really nice paper. A few small suggestions for additions:

Q.i) Task. As the authors note, the task is deliberately simple, which is appropriate given the analytical treatment. Do the authors have intuitions about how tasks requiring multiple timescales, or a readout time that is not fixed, would behave under these constraints? The fixed readout time seems to particularly favor transient solutions, and the general formulation in App. A.1 already accommodates a distribution over readout times. I am not requesting new experiments, but a slightly expanded discussion would be interesting.

Q.ii) Biological interpretation: Again, as mentioned in the discussion, penalizing absolute activity rather than deviations from baseline may reduce biological plausibility. This seems most relevant for the ALM model, where the cost acts on firing rates. Since that model already has trainable biases, retraining with a baseline-deviation penalty seems relatively straightforward and would make an informative supplementary section.

2 comments

We thank the reviewer for their thoughtful comments, and are glad to hear that they liked our work. We address the points raised under Questions below:

How results would change with a more complex task:

As the reviewer notes, there are straightforward ways of extending our analytical framework to incorporate a distribution over readout times rather than a pointwise constraint on the network output, as discussed in the Appendix. However, the same level of analytical insight becomes significantly harder to achieve. We had at one point attempted to determine whether a nonzero variance of the readout time distribution could be interpreted as a rescaling of the effective λ\lambda of the optimization problem, but were unable to make much analytical progress on this question. We generally agree with the reviewer in that we suspect that more uncertainty in the readout time should increase pressure for solutions to behave more like the persistent solution that is selected by the weight efficiency objective; this would be consistent with Stroud et al. 2024. That said, the observed tradeoff between persistent and transient solutions with λ\lambda should still persist even when the output constraint is more complex (e.g. continuous rather than pointwise), as was demonstrated numerically in the nonlinear RNN experiments in Fig. 4.

Biological interpretation:

We agree with the reviewer that investigating how alternative formulations of the objective costs studied here might affect the solutions that are selected is worthwhile. However, we prefer to focus the presentation of this current paper on the simplest version of these objectives, so that the number of moving knobs (here, just λ\lambda and γ\gamma/noise level) remains relatively controlled and therefore easier to interpret. We will leave a more detailed investigation into how different objectives (baselined activity, sparsity instead of squared norm, etc.) affect the properties of optimal solutions for future work.

I have already given the maximum score. I think the paper should be accepted. Nothing further is needed from my view.

Reviewer_2dwfOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper tackles the problem of exploring the solution space of a simple recall task, first in linear RNNs and then in non-linear RNNs. The linear RNNs are tackled analytically with a neat mathematical formulation. The paper's key result is that, under certain conditions, robustness to noise requires a trade-off against activity efficiency (measured as L2 norm of the activity), and that this trade-off might explain some phenomena observed in the mouse anterior lateral motor cortex.

Critical strengths
  1. The analytical solution for the linear RNN is neat and relatively straightforward. The problem is overall treated with good mathematical rigour and explained reasonably well.
  2. The paper contains a good number of results that are well presented and tied into an interesting narrative.
  3. Results on non-linear RNNs and neuroscience findings add richness to the paper.
Critical weaknesses
  1. Dependence on absolute noise scale. Perhaps the most important finding of the paper (the trade-off between robustness and efficiency) holds only when considering absolute noise. But when considering signal-to-noise ratio (as in Ganguli 2008), makes the combination of these two actually optimal (!). The paper's justification for picking this is somewhat circular and dependent on this particular choice of task.
  2. Novelty is limited. Although some results are novel, and it is still useful to see all these results synthesised nicely, many of these findings have been reported before in very similar setups.
  3. The paper does not engage sufficiently with prior work. It's mentioned quite briefly in the introduction and discussion, but connections are not drawn with sufficient depth. Some glaring omissions from the reference list include e.g. "Aligned and oblique dynamics in RNNs" (Schuessler, 2024) and "Improved memory in RNNs with sequential non-normal dynamics" (Orhan 2020).
  4. The paper is at times hard to follow, and it could do a better job at explaining its core concepts. For example, the concept of "misaligned coding" comes up multiple times, but it is not described explicitly and in sufficient detail.
Questions for authors

Please reply to weakness number 1. Why is the absolute variance the right measure to optimise here? Why wouldn't e.g. the anterior lateral motor cortex optimise SNR instead?

3 comments
AuthorsOpenReview ↗

Response Part 1

We thank the reviewer for their balanced appraisal of our work, and appreciate their overall positive assessment. We address each point below:

Dependence on absolute noise scale:

As the reviewer notes, this stipulation about absolute noise scale, and how the results might change under different objectives like SNR, are already made clear in the text. We also gesture towards this broader limitation of normative theories in general in the Discussion (i.e. why use this objective and not that). That said, we argue that our justification for focusing on absolute noise sensitivity is not circular. Indeed, optimizing SNR without retaining the input-output gain constraint would introduce additional scaling freedoms, allowing the network to improve SNR without actually recovering the scale of the observed stimulus. Since the purpose of our task is precisely to recover the scale of the stimulus, we fix this gain to 11, under which minimizing absolute output-noise variance is equivalent to maximizing SNR (we make note of this in the main text now). We understand why this may feel like an unsatisfactory justification, but analytically addressing a more SNR-centric version of this problem would require an entirely different task and formulation, and thus is out of the scope of this paper. Furthermore, if one wishes to use task-trained RNNs as hypothesis generators for neuroscience, then the effective penalty incurred due to noise sensitivity is much more directly related to this adjoint cost than SNR (see the discussion around this point in the main text); this alone makes it valuable to understand the competition between activity efficiency and absolute noise sensitivity.

Regarding Ganguli et al. 2008, it is not the case that the SNR-optimal networks discussed there necessarily achieve both high activity efficiency and noise robustness. Rather, the SNR-optimal networks they construct rely on strong transient amplification of both signal and noise due to underlying feedforward structure, with signal activity and noise variance both growing (exponentially in some cases) as the signal propagates through the network. In other words, high SNR in these networks does not imply simultaneous activity efficiency and robustness, and is only achievable because the signal is strongly amplified to match the exploding noise scale. This is entirely consistent with our result.

Most importantly however, there is a fairly direct way to massage the tradeoff inequality in Eq. (6) into one relating SNR to activity cost. If the gain g(τ)g(\tau) is allowed to be optimized as well (otherwise SNR is effectively identical to inverse absolute noise), then we straightforwardly obtain by similar arguments that ExEq(τ)≥g(τ)2τ2E_x E_q^{(\tau)} \geq g(\tau)^2\tau^2, where Eq(τ)=∫0τ∥q(t)∥2dtE_q^{(\tau)} = \int_0^\tau \|\mathbf{q}(t)\|^2 dt, and thus can upper bound SNR(τ)=g(τ)2ση2Eq(τ)\text{SNR}(\tau) = \frac{g(\tau)^2}{\sigma^2_\eta E_q^{(\tau)}} as SNR(τ)≤1ση2τ2Ex\text{SNR}(\tau) \leq \frac{1}{\sigma^2_\eta \tau^2}E_x. Morally, this is expressing a similar fundamental constraint as what we state: to achieve high SNR, one needs to pay in activity cost.

More broadly, we would also like to clarify that we do not view this tradeoff between efficiency and robustness as the most important result of the paper. This is certainly a neat result that ties together all the desiderata for biological circuits that we consider. However, as the framing in the Introduction, Discussion, and choice of nonlinear RNN experiments demonstrate, our main result is the dissociation between explained variance and causal influence emerging in the activity-constrained, rather than weight-constrained regime, and its normative implications. In particular, the choice of noise metric is not central to this result; it is only relevant to our ALM experiments insofar as injected noise during training was necessary to produce activity consistent with experimentally observed behavior during catch trials.

AuthorsOpenReview ↗

Response Part 2

Concerns about limited novelty and connections with prior work:

Regarding engaging sufficiently with prior work, our discussion section was on the terse side due to page limits; we have expanded it in the revised version. We apologize for missing Schuessler et al and Orhan 2020 initially; both citations have been added now. The expanded discussion section now also reflects the relationship between concepts studied in this work and those studied in Schuessler et al (see below), as well as further expounding upon how our study relates to Ritter and Chadwick.

Regarding limited novelty, we invite the reviewer to clarify which results presented here they believe to be not particularly novel. We speculate that they are referring to the fact that the role of nonnormal/transient dynamics in efficient memory maintainence has already been the subject of some prior studies. We acknowledge this, but we believe we sharpen this point considerably by 1) connecting these ideas to an underlying normative objective of activity efficiency, with fairly detailed analytics characterizing when, and to what extent the optimal network emulates a long chain 2) contrasting this with the dynamical motifs evoked by other normative objectives relevant to both the theory and practice of RNNs as hypothesis generators for neuroscience, and 3) drawing novel connections to recently identified phenomenology in premotor circuits. To elaborate on the point about analytic results, many prior works in this area have been primarily numerical/empirical in nature (e.g. Orhan 2020), or have only considered restricted versions of similar optimization problems where only the input weights are optimized, for fixed recurrent weights (e.g., the theoretical analysis of Stroud et al. 2023). To address the relation of our work with Schuessler et al 2024 specifically, there is a fairly sharp mathematical distinction between the oblique dynamics discussed there and the phenomenon of misaligned coding focused on here (although the concepts are related). Schuessler et al. focus on the alignment (or lack thereof) of the dominant modes of activity with the readout direction c\mathbf{c}, which is a primarily geometric notion of misalignment. In our setting, misaligned coding refers to dominant modes of activity having little causal influence on the readout direction at the relevant future time (in this case, τ\tau), so that the relevant vector we consider alignment with is not the readout direction itself, but the time-evolving adjoint q(τ−t)=eA⊤(τ−t)c\mathbf{q}(\tau-t) = e^{\mathbf{A}^\top (\tau-t)}\mathbf{c}. We have now made note of this distinction in the Discussion.

Parts of the paper being hard to follow:

We are sorry to hear this and would like to know if there are any other specific points of confusion we can help clarify. On the particular point about "misaligned coding": we introduce this term in the introduction directly after discussing the dissociation between modes of high explained variance and modes of high causal influence observed in ALM. A very similar clarification is made at the end of section 2.2. We believe this is already sufficiently explicit of a definition, and would stylistically prefer to keep the current wording choice rather than belaboring the point. In addition, our expanded discussion of how this concept compares to the oblique dynamics explored in Schuessler et al. should clarify the term even further. However, if the reviewer still feels strongly that this term is still not clear, we would be happy to amend our introduction accordingly.

Reviewer_2dwfOpenReview ↗

Reviewer acknowledgement

Thanks to the authors for their response, the clarification re. SNR is particularly helpful. I still think the paper could be improved in terms of readability to be more accessible for readers, but nonetheless I find this paper makes a good contribution and I'm happy to recommend it for publication.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.