acceptodds
Under review as a conference paper at ICLR 2027

An efficient coding account of representational drift

Abstract

Neural responses to stimuli change continually over days even as perception and behavior remain stable, a phenomenon called representational drift. Drift has been modeled with unsupervised learning objectives that admit a degenerate solution space of representations that preserve input similarity structure. But such objectives are ill-suited for early sensory areas, where the code must allocate sensitivity according to input statistics rather than reproduce similarity structure. Here, we test the hypothesis that drift in these areas instead arises from mechanisms supporting efficient coding, formalized as maximizing Fisher information about the stimulus in the neural population. We show that a biologically plausible uncertainty-gated differential Hebbian plasticity rule, in which synaptic change depends on the time derivatives of pre- and post-synaptic activity, implements gradient ascent on Fisher information, and that adding synaptic noise produces substantial drift in individual neurons while population Fisher information remains stable. We confirm this prediction in longitudinal calcium imaging from the primary visual and auditory cortex in mice, where single-neuron tuning drifts over days while population Fisher information does not. These findings offer a novel, efficient coding account of drift in sensory cortex.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Fisher information, efficient coding, representational drift

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qTBqOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a theoretical account of representational drift in early sensory cortex. Starting with an efficient coding objective (inspired by Ganguli and Simoncelli) they derive differential Hebbian upate rules that perform stochastic gradient ascent on the expected log Fisher information. The update rule depends on pre- and post-synaptic activation as well as an uncertainty-dependent signal that varies per stimulus, U(s), and additive Gaussian noise. They compare this rule to existing similarity matching objectives that also claim to account for representational drift. The authors show that single neuron tuning drifts, while the population fisher information (FI) profile stays constant over time. The similarity matching network (SM) also drifts but has an uncorrelated FI profile. In both cases, the RSM is preserved, complicating its use as a signature of theoretical models of drift. The authors then analyze existing data from auditory and visual cortex in mice and report drifting tuning with stable FI and RSMs, which they claim supports efficient coding over SM models of drift.

Critical strengths

The paper overall is clearly written and easy to follow. The framework is experimentally testable and fits in well with existing efficient coding models. The three-factor update rule is normative and the time-derivative version adequately addresses the issue of locality for the pre and post derivatives. The authors also validate their findings in nonlinear networks and with time derivatives, strengthening their derivations. The data analysis done is across two sensory modalities which also strengthens the broad applicability of their work, and the limitations of the work are well addressed.

Critical weaknesses

1.The data analysis doesn't adequately establish that the efficient coding/differential Hebbian rule are the source of drift. In part because the theory/simulation design doesn't closely match the data. For example, the equation on line 163 imples that static stimuli produce no learning at all, since the weight updates are driven by changes in the stimulus through v^2. The data considered are from experiments with pure tones or fixed direction gratings that are static in the encoded dimension of interest, so it doesn't seem like this update rule would account for the mechanism in the recordings. The observed data in the drift could instead come from habituation, behavioral state, etc. Moreover, the data analysis could be more airtight. The metric the authors use to measure the FI stability is the pearson correlation, which is blind to changes in amplitude, and the data do clearly show changes in amplitude over time (see fig. 3D especially). The authors don't address this instead saying that FI is conserved over days. The model's FI also decays slightly, but the source of this theoretical decay isn't discussed or linked to the decay in the data. Moreover, the population vector (in)stability isn't shown, which is the most commonly used metric of representational drift in populations to my understanding. Cross-day decoding is also absent for the data and the model. Finally, the paper does not adequately rule out similarity matching in terms of the FI metric. The correlation in FI for the SM model being ~ 0 does not show that similarity matching fails to preserve FI. The authors note that the FI of the output is inherited from the input for this model (section 2.5) The input units have identical tuning so the input FI should be approximately flat, modulo some noise. If this is the case then you would expect that the pearson correlation of two flat lines should be ~ 0. The authors therefore fail to show whether this is inherited from the input or the SM itself. The data are clearly non zero, but in the brain the input would almost certainly not be flat.

  1. The claims in the paper are too strong for what is actually shown, and should be toned down. The paper claims to have a novel normative theory of drift, but the drift mechanism is noise driven diffusion on a solution manifold, which has been discussed in previous work by Pashakhanloo & Koulakov, 2023. The claims in this paper would be significantly strenghtened if the differences between this work and Pashakhanloo & Koulakov were discussed. The normative contribution of the paper instead seems to be in the learning rule; the authors, however, do not establish whether the drift is driven by this learning rule or by the additive noise in the learning rule. For example, does drift still occur if \sigma_w = 0? The paper spends considerable time discussing that the model results show that the efficient coding perspective and learning rule explain early sensory coding over the similarity matching objective. However, the results in Qin et al 2023 are meant to describe drift in the PPC and hippocampus, not in sensory areas, making it seem a bit like a strawman argument; showing that SM fails in auditory and visual areas is interesting, but these claims should be more measured. Moreover, drift in the olfactory cortex, which is quite well studied and even cited in the paper introduction (Schoonover et al., 2021), is not addressed here. It is not clear how the model would account for non-continuous stimuli. Also, one of the central results in Schoonover et al., 2021 is that more frequent stimuli drift more slowly in the model -- this is a hallmark of representational drift that is not addressed in the paper, and one which an overarching theory of drift in early sensory areas should address. Finally, the plausibility of the U(s) signal is overstated. The cited papers don't show that ACh fluctuates in the stimulus-specific trial-by-trial way that the theory requires, in particular during passive exposure.
Questions for authors
  1. A.2 assumes that the learned solution W has a fixed peak and w \propto 1/d (like in Ganguli an dSimoncelli) but this is not checked. Please check that width is proportional to stimulus density.
  2. How does the Gaussian additive noise assumption change the results from the Ganguli and Simoncelli paper, which used Poisson noise? Do the results from the Gaussian derivation hold for Poisson noise?
  3. Inputs are assumed to be noiseless, which is not biologically plausible even for early sensory areas. How does input noise affect results?
  4. The model relies on a learning signal U(s) which in turn requires current knowledge of the drifting tuning curves that is assumed to be calculated downstream. This is a circular assumption, given that tracking this tuning is precisely the problem of drift. Could you discuss this, and show robustness to U(s) which is either lagged or bias or inexact?
  5. The simulations rely on \dot{y} which is ill defined for small \delta t - this should be noted.
  6. Certain hyperparameters are missing (\sigma_w, SM learnign rates).
  7. Typos in the text often show e.g. "equation equation 3" (line 190).
  8. Are the drift rates between the SM and efficient coding models matched? It would be good to re-run simulations with matched drift rates s/t the RSM and FI changes are on more equal footing.
  9. What is the dependence of the model on M and N? M > N necessarily adds degeneracy. Does drift increase as M increases relative to N?
  10. The simulations rely on continuous and smoothly varying stimuli. How would this model apply to non-continuous sensory input (like in the olfactory cortex)? Moreover, the experiment used in the auditory dataset has different structure from the simulations (frequency is bounded, not circular) - could you discuss how this may affect your results?
  11. Does drift occur when \sigma_w is zero? How do drift and FI depend on \sigma_w? How does this account relate to the SGD-driven drift in Pashakhanloo & Koulakov, 2023?
  12. Do the population vector correlations and cross-day decoding show readout degradation despite stable FI in the data and the model?
  13. Suggest renaming v^2 to avoid confusion with v_t.
  14. V1 day ranges are different between fig 3 and fig 4. Could you add the later days to figure 3?
  15. Is there evidence for the environmental dependence of tuning width in early sensory cortex to motivate using the Ganguli and Simoncelli framework? If so, please cite, if not, please acknowledge.
  16. RSM changes are measured by Euclidean distance, which is not sensitive to noise. Could you also show how e.g. the Mahalanobis distance changes over learning?
  17. The exponent fit of 2 is only tested with one prior. Does that change if the prior is e.g. more sharply peaked?
3 comments
AuthorsOpenReview ↗

Response to reviewer qTBq

We thank the reviewer for detailed feedback. Below, we address the points not adequately covered in the general rebuttal (W: Weakness, Q: Question).

W1. Model/data mismatch. We agree the rule requires temporal variability in stimuli, and the Methods now state that we model natural sensory input rather than discrete stimuli. We interpret between-session drift as ongoing plasticity driven by natural experience. Yes, habituation and behavioural state may also contribute to drift but we focus on what population quantities remain stable despite drift and what plasticity rules might be driving that. We apologize this was not clear.

W2. Amplitude-blind metric. Please see general response 2 - we repeats the analysis with a scale-sensitive metric (Figure A22). The model’s apparent slight FI decay was an artifact of the comparison window whose earliest time point preceded learning convergence. It no longer decays when comparing time points after convergence.

W3. Population vectors and cross-day decoding. Both were found to decay and now reported including the corresponding ceiling/chance levels (Figure A14).

W4. Flat input FI. Thank you for this observation, please see general response 1. We now use heterogeneous input tuning for isotropic noise yielding a U-shaped FI profile (Figure A17), and rerun model comparisons with correlated noise (Figure 4).

W5. Strong claim about novel normative theory of drift. Please see general response 4. Drift occurs without synaptic noise, but slowly (Figure A6). We agree noise-driven diffusion on a solution manifold is not new, and the Discussion compares to Pashakhanloo and Koulakov.

W6. Qin et al. concerns PPC/hippocampus. The Discussion now states this and frames our focus as the solution space in the sensory cortex. We also state that our comparisons do not rule out similarity matching and suggest experiments to distinguish them.

W7. Olfactory cortex and non-continuous stimuli. Drift is slower for more frequent stimuli in our simulations (Figure A5). The Discussion now states that FI requires a continuous variable, so generalizing to olfaction would need a rule optimizing mutual information directly.

W8. U(s) plausibility overstated. We agree and have replaced biological plausibility with local in the abstract. Acetylcholine is presented as a candidate, and the Discussion states that U is computed from the normalized likelihood, which a downstream area might not have direct access to.

Q1. The width assumption. Thank you for this important question. The learned widths were uniform, so the width assumption of Ganguli and Simoncelli does not hold in our model. The weight budget constrains the integral of J, so the rederived optimum is J* proportional to p rather than p squared (Appendix A.2). The exponent of 2 in the original simulations arose from input tuning that was not sufficiently narrow which biases the Fisher information profile in small networks, and we mistook that the widths depended on density without checking. We apologize for the oversight. With sufficiently narrow input tuning, the exponent measured in our revised simulations is around 1 (Figure A2).

Q2. Poisson noise. Results hold for Poisson noise (Figure A4) because the optimum depends on the resource constraint which does not alter the response amplitude.

Q3. Input noise. Results hold when inputs have nominal noise (Figure A7).

Q4. Inexact U(s). Results are robust to biased/noisy U, but not to delay i.e., when U is estimated using drifted tuning curves which we discuss in the text (Figure A8).

Q5. Time derivative at small step sizes. A variant accumulating the derivative product in an eligibility trace removes the dependence on step size and reaches the same FI allocation (Appendix A.5, Figure A3).

Q6, Q7, Q13. We added hyperparameter values, fixed equation references, and velocity symbol.

Q8. Matched drift rates. All comparisons are now done with models calibrated to comparable drift rates (Figure A16).

Q9. M and N. FI profile and stability hold for M/N from 0.2 to 10 (Figure A9). Drift rate was independent of M because synaptic noise is N-dimensional for each neuron.

Q10. Bounded auditory axis. The model applies to bounded variables. For data, results were similar when ignoring the extreme tones or using one-sided differences at the extreme tones, so circularity was not assumed.

Q11. Please see W5 response.

Q12. Please see W3 response.

Q14. Data day ranges are now consistent across figures.

Q15. Evidence for environmental dependence of tuning width. As discussed in response to Q1, our findings do not hinge on that assumption.

Q16. Mahalanobis distance. Stability is comparable under both euclidean and Mahalanobis metrics (Figure A22) because the FI profile is exactly the near-diagonal band of the Mahalanobis matrix.

Q17. Sharper prior. FI profile and exponent hold across prior sharpness and network sizes (Figure A2).

AuthorsOpenReview ↗

Checking in on our rebuttal

Thank you again for your helpful feedback. We wanted to check if you have had a chance to look at our revisions, and if there are any remaining questions or clarifications we can provide before the discussion period closes.

AuthorsOpenReview ↗

Checking in again

Dear reviewer qTBq, we wanted to check in once more since the discussion period ends today. We made extensive revisions to address your comments including a correction to the derivation in appendix A.2. Thanks again for your detailed feedback - we found them extremely helpful and we'd appreciate if you could take a look at the revisions and responses and let us know if there are pending concerns.

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

The paper proposes that efficient coding with noisy synaptic updates causes representational drift in sensory cortex. Efficient coding is formalized as maximizing expected log Fisher information about a stimulus. This maximization is implemented as an uncertainty-gated differential Hebbian update. The authors present simulations and analyses of auditory and visual cortex data. They argue that the observed representational drift in both is in line with efficient coding as opposed to similarity matching.

Critical strengths

Relating representational drift to efficient coding by deriving an uncertainty-gated learning rule from Fisher information is novel and valuable. Moreover, hypotheses were extensively tested by using simulation and two datasets from different sensory modalities.

Critical weaknesses

The major concerns are with logic and hypothesis testing. First, stable representational similarity implies stable Fisher. J(s) is determined by distances between responses to neighboring stimuli, which are also used as entries in RSA. Similarity matching then also predicts a stable J profile, so observing stability here does by definition not favor efficient coding. In the results, similarity is at least as stable as J in both areas, consistent with similarity matching. Second, the main finding of the paper, that stability of FI under similarity matching is zero while high for the efficient coding model is uncontrolled, and thus unconvincing

Questions for authors

Distinguish clarification requests from suggestions for future work.)

  1. Important controls are missing throughout, making findings difficult to interpret. For example, the FI similarity matching results in 4B could be due to the uniform input, can the authors test this?
  2. Can the authors specify the similarity-matching network better? Equation 1 does not apply when the noise output is shaped by lateral dynamics.
4 comments
AuthorsOpenReview ↗

Response to reviewer SXGs

We thank the reviewer for clear feedback. Below, we address the points not adequately covered in the general rebuttal (W: Weakness, Q: Question).

W1. Stable representational similarity implies stable Fisher information, so the observation does not favour efficient coding.
We thank the reviewer for noting this. Please see general response 1. The identity is now stated in Methods section 2.5 and in the Appendix A.3. The statement holds for isotropic noise. The Fisher information in the recordings is estimated using each session’s measured residual covariance, and for general Σ the Fisher information profile is the near-diagonal band of the Mahalanobis rather than the Euclidean dissimilarity matrix, so conserved pairwise distances do not imply a conserved profile when the noise is correlated. Our updated model simulations with correlated noise validate this theoretical possibility (Figure 4A,B).

But we do agree that representational similarity is at least as stable as the Fisher information profile in both areas and state explicitly in the revision that these observations do not falsify the similarity matching account. Carefully arbitrating between the two accounts would require manipulating stimulus distributions and recording from input and output neural populations, and our work makes falsifiable predictions for such experiments as we note in the discussion.

W2. The main finding, that Fisher information stability is zero under similarity matching and high under efficient coding, is uncontrolled.
We agree that for the comparison that used a uniformly tiled input and isotropic noise, the near-zero correlation reflected the flat input profile rather than the objective. The comparison is now run with the same correlated noise in both models and with input statistics and drift rates matched (Figure A16). We additionally tested a variant of the similarity-matching network that enforced non-negative output response (Appendix A.6, Figure A18). Representational similarity is conserved in both models, and the Fisher information profile is conserved in the efficient coding model and not in the similarity-matching model. The heterogeneous-input control under isotropic noise is reported separately to clarify that in that case, similarity matching model conserves both measures (Figure A17).

Q1. The Fisher information result for similarity matching in Figure 4B could be due to the uniform input. Can the authors test this?
Yes, it was. For the uniformly tiled input, the input Fisher information profile is flat and the similarity-matching output inherits it, and a correlation between two noisy flat lines is uninformative as reviewer qTBq also pointed out. We tested heterogeneous input tuning, under which the profile is inherited and non-flat (Figure A17), and the revised comparison in Figure 4A,B instead uses correlated noise to separate the two accounts.

On controls more generally, general responses 2 and 3 list the analyses added to the revision, which include the within-session reliability ceiling and the across-animal floor for the recordings, and robustness simulations for the model spanning the output nonlinearity, the noise model, input noise, network size, prior sharpness, M relative to N, and the accuracy of the uncertainty signal.

Q2. Can the authors specify the similarity-matching network better? Equation 1 does not apply when the output noise is shaped by lateral dynamics.
We apologize for omitting this detail. We simulated two variants of the similarity-matching network. The linear variant is now specified in Methods section 2.5 and shown in Figure 4. Briefly, we simulate the linear Hebbian/anti-Hebbian network of Qin et al. (2023), whose steady-state response is μ(s) = L⁻¹Wx(s), with drift arising from noise in the weight updates L (lateral) and W (feedforward). Responses in that network are deterministic given the weights, so the Fisher information is not defined without a model of trial-to-trial variability. Because Fisher information is invariant under any invertible linear transformation applied to both the mean response and its noise, noise injected upstream of the lateral stage would leave the profile unchanged by the lateral weights. We therefore add output noise, y = μ(s) + ξ with ξ ~ N(0,Σ), with the same covariance Σ used in the efficient coding model. For stricter comparison with the efficient coding model, a variant of the network that produces non-negative outputs is described in Appendix A.6 and corresponding simulation results shown in Figure A18.

AuthorsOpenReview ↗

Checking in on our rebuttal

Thank you again for your helpful feedback. We wanted to check if you have had a chance to look at our revisions, and if there are any remaining questions or clarifications we can provide before the discussion period closes.

AuthorsOpenReview ↗

Checking in again

Dear reviewer SXGs, we wanted to check in once more as the discussion period ends today. Thanks again for your feedback - we believe we have addressed your primary concern. Briefly, we reran the comparison with the similarity-matching model under settings where it is dissociable from efficient coding (correlated noise and heterogeneous input tuning). We would appreciate it if you could take a few minutes to look at the revisions and let us know if there are pending concerns.

Dear authors, I appreciate the reviews and I think that they do improve the results. However, given the responses I am inclined to maintain my score, as the initial claim. The correlated noise statement seems a significant change that would need a full reassessment of the paper. As a personal note, I think that the paper might be better fit for a neuroscience journal with a more slow review process.

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

This paper proposes an efficient coding account of representational drift in early sensory cortex. Motivated by the observation that individual neuronal tuning properties change over time while population-level representations remain relatively stable, the authors argue that efficient coding may provide a more appropriate normative explanation for representational drift in sensory regions than previously proposed similarity-matching objectives.

The authors formulate efficient coding as the maximization of expected log Fisher information under synaptic resource constraints. They derive an uncertainty-gated differential Hebbian plasticity rule that approximately implements gradient ascent on this objective. By introducing stochastic synaptic updates, they demonstrate that individual neuronal tuning curves can drift substantially while the population-level Fisher information profile remains stable.

The theoretical predictions are evaluated through numerical simulations and longitudinal calcium-imaging datasets from mouse primary visual and auditory cortex. The empirical analyses show that population Fisher information remains relatively stable across recording sessions despite substantial changes in individual neuronal responses. Comparisons with a similarity-matching model further suggest that the proposed efficient coding framework better reproduces the observed preservation of Fisher information.

Overall, the paper presents a theoretically motivated account of representational drift and provides empirical observations consistent with its predictions. However, the extent to which these findings establish the proposed mechanism, rather than demonstrate compatibility with an existing neural coding principle, remains unclear.

Critical strengths

S1. A theoretically coherent connection between efficient coding and representational drift

The strongest contribution of this work is its attempt to connect representational drift to a normative efficient coding objective through an explicit synaptic plasticity mechanism.

Rather than treating representational drift solely as stochastic movement within a degenerate representational space, the authors derive an uncertainty-gated differential Hebbian learning rule from the maximization of Fisher information. This provides a mathematically motivated relationship between a population-level coding objective, local synaptic plasticity, and the emergence of representational drift.

Although the individual components, including efficient coding, Fisher information maximization, and differential Hebbian plasticity, have substantial precedents, their integration into a unified account of representational drift is conceptually interesting.

This theoretical coherence is a meaningful strength, particularly for understanding how stable population-level coding properties might coexist with continuously changing neuronal representations.

S2. Empirical evaluation using longitudinal neural recordings

Another strength is the evaluation of the theoretical predictions using longitudinal calcium-imaging datasets from both mouse visual and auditory cortex.

The observation that population Fisher information remains relatively stable despite substantial single-neuron tuning drift provides empirical support for an important prediction of the proposed framework.

Furthermore, the comparison with a similarity-matching model attempts to distinguish between competing normative accounts rather than merely demonstrating that the proposed model can reproduce representational drift.

Although the evidence does not conclusively identify the underlying mechanism, the combination of theoretical derivation, numerical simulations, and analyses of experimental neural recordings makes the study more substantial than a purely conceptual proposal.

Critical weaknesses

W1. The empirical evidence does not establish the proposed efficient coding mechanism

My primary concern is the gap between the proposed mechanistic explanation and the evidence presented to support it.

The central empirical observation is that population Fisher information remains relatively stable across recording sessions despite changes in individual neuronal tuning. While this observation is consistent with the proposed efficient coding framework, it does not establish that the observed representational drift is driven by optimization of the efficient coding objective.

In particular, the preservation of Fisher information is not necessarily unique to the proposed mechanism. Alternative learning objectives or constraints, including homeostatic regulation and other information-preserving mechanisms, could potentially produce similar population-level stability.

The comparison with similarity matching is informative but insufficient to establish mechanistic specificity. Demonstrating that one particular implementation of similarity matching fails to preserve Fisher information does not rule out other mechanisms that may reproduce the same empirical observations.

Furthermore, the authors acknowledge that their experiments do not directly demonstrate Fisher information maximization in the biological system. The neural recordings therefore provide evidence of consistency with the proposed account rather than direct validation of the underlying optimization principle.

This distinction is consequential because the primary contribution of the paper is framed as a mechanistic explanation of representational drift. Without stronger evidence distinguishing the proposed mechanism from plausible alternatives, the central scientific conclusion remains underdetermined.

W2. Limited methodological contribution and unclear implications for machine learning

Although the proposed framework is interesting from a computational neuroscience perspective, its methodological contribution to the broader machine learning community is not sufficiently established.

The paper combines established concepts, including efficient coding, Fisher information maximization, differential Hebbian plasticity, and stochastic synaptic dynamics, to explain a specific neuroscientific phenomenon. The derivation of the uncertainty-gated learning rule is theoretically meaningful, but the manuscript does not demonstrate whether this formulation provides a generalizable learning principle beyond the particular sensory coding setting considered.

For example, it remains unclear whether the proposed mechanism offers advantages for representation learning, continual learning, robustness to representational drift, or other machine learning problems in which stable information must be maintained despite changes in internal representations.

The experimental validation is also restricted to reproducing neural observations, without investigating whether the proposed objective or learning rule leads to meaningful computational benefits relative to existing learning approaches.

I do not consider the absence of conventional machine learning benchmarks inherently problematic for a theoretically oriented submission. However, the paper should establish either a sufficiently general methodological advance or a broader theoretical insight relevant to machine learning.

In its current form, the work reads primarily as a computational neuroscience study proposing an alternative normative explanation for a biological phenomenon, rather than as a contribution that substantially advances machine learning methodology or theory.

Questions for authors

Q1. What evidence distinguishes efficient coding from alternative mechanisms that preserve population Fisher information?

The primary empirical finding is that population Fisher information remains stable despite representational drift. However, it is unclear whether this property is specific to efficient coding or could emerge from other learning objectives and constraints.

Could the authors clarify which empirical predictions uniquely distinguish the proposed efficient coding mechanism from alternative explanations, beyond the particular similarity-matching implementation considered?

In particular, are there additional predictions of the proposed model that can be evaluated using the existing longitudinal recordings and that would help establish mechanistic specificity?

This question is central to my assessment, as the proposed mechanism constitutes the main scientific contribution.

Q2. Are efficient coding and similarity matching necessarily competing explanations?

The manuscript positions efficient coding as an alternative to similarity matching, particularly for early sensory cortex.

However, the two objectives are not necessarily mutually exclusive. Indeed, the proposed efficient coding model also exhibits relatively stable representational similarity, suggesting that the preservation of representational geometry may emerge under multiple objectives.

Could the authors clarify whether their theoretical framework establishes any fundamental incompatibility between efficient coding and similarity matching?

More importantly, does the proposed framework provide predictions that distinguish efficient coding from models optimizing both objectives simultaneously?

Without such clarification, it is difficult to determine whether the proposed account replaces an existing explanation or simply introduces another compatible objective.

Q3. How robust are the empirical Fisher information estimates to covariance estimation and dimensionality reduction?

The empirical Fisher information is estimated using the inverse residual covariance matrix, with covariance inversion performed in a PCA subspace explaining 90% of the residual variance.

Given the limited number of trials relative to the number of recorded neurons, the estimated Fisher information may be sensitive to covariance regularization, dimensionality reduction, and sampling variability.

Could the authors provide sensitivity analyses examining whether the reported cross-session stability of Fisher information persists under different PCA variance thresholds or covariance estimation procedures?

It would also be useful to establish whether the observed stability exceeds the reliability expected from repeated estimates within the same recording session.

These analyses would help determine whether the reported stability reflects a robust property of the neural population rather than characteristics of the estimation procedure.

Q4. How does the proposed mechanism generalize beyond the simplified assumptions used in the theoretical model?

The theoretical derivation assumes a relatively simple feedforward architecture with additive independent Gaussian noise and constrained nonnegative synaptic weights.

However, the empirical Fisher information analysis explicitly incorporates correlated neural variability, and biological sensory representations involve considerably more complex dynamics.

Could the authors clarify whether the central prediction of stable population Fisher information under representational drift continues to hold when noise correlations are introduced into the learning dynamics?

More specifically, is the preservation of Fisher information a general consequence of the proposed learning objective, or does it depend critically on the particular assumptions and constraints used in the simulations?

Q5. What is the broader machine learning implication of the proposed learning rule?

The derivation of the uncertainty-gated differential Hebbian rule is interesting, but its implications beyond the specific neuroscientific setting remain unclear.

Could the authors clarify what generalizable insight the proposed mechanism provides for machine learning, particularly for representation learning or continual learning?

For instance, does the proposed objective offer a principled way to maintain task-relevant information while allowing internal representations to change, and is this property meaningfully different from what existing representation-learning objectives already provide?

I am not necessarily requesting additional large-scale benchmarks. Rather, a clearer theoretical or computational argument establishing the broader relevance of the proposed mechanism would materially affect my assessment of the paper’s contribution to ICLR.

4 comments
AuthorsOpenReview ↗

Response to reviewer Vbib

We thank the reviewer for clearly identifying the strengths and limitations. Below, we address the points not adequately covered in the general rebuttal (W: Weakness, Q: Question).

W1. The empirical evidence does not establish the proposed efficient coding mechanism.
We have extended the model comparison to a nonnegative variant of the similarity-matching network in which the output activity is sign-constrained similar to the efficient coding model but we agree this does not establish mechanistic specificity. We have gone through the manuscript to ensure the recordings are not presented as evidence for efficient coding, and that the data are described throughout as consistent with the model. This limitation is stated explicitly in the Discussion, as the reviewer noted, and that point is now expanded to state that other population-level mechanisms, including homeostatic regulation, could in principle preserve Fisher information.

Our primary contribution is to connect the classical theory of efficient coding in sensory cortex with the recently documented phenomenon of representational drift, to derive a falsifiable population-level prediction, that the Fisher information profile is conserved while single-neuron tuning changes, and to provide a plausible learning rule that implements the objective. Identifying a candidate objective is informative in its own right, since it specifies which population quantity should be conserved, and the consistency of the existing recordings with that prediction is a step toward experiments that test more specific predictions about the mechanistic aspects of the model.

Lines 506-521 of the discussion specifies what those predictions are. Definitively separating the candidate objectives requires tracking the same neurons across days while manipulating the stimulus statistics in highly controlled environments. A plausible implementation together with falsifiable predictions provides the basis for designing such challenging experiments.

W2. Limited methodological contribution and unclear implications for machine learning.
We believe there are historical precedents for work like ours to appear under the area of applications to neuroscience & cognitive science. In the revised version, we discuss a link to continual learning. Please see the last paragraph of the discussion (lines 533-539).

Q1. What evidence distinguishes efficient coding from alternative mechanisms that preserve population Fisher information?
Please see our response to W1 above and general response 1. Conservation of the Fisher information profile on its own does not distinguish the accounts. What distinguishes them is whether the profile depends on the stimulus distribution. Efficient coding predicts that it reallocates when the distribution changes, whereas a profile inherited from the input or maintained homeostatically does not. The present longitudinal recordings cannot address this, since the stimulus distribution is not manipulated. The predictions in the Discussion explicitly identify such candidate manipulations including blocking the uncertainty signal that gates plasticity. What our model comparison does establish is that among the objectives we considered, similarity matching conserves representational similarity but not the Fisher information profile, suggesting that an account based on similarity alone is incomplete. Please also see our response to W1 above and general response 1.

Q2. Are efficient coding and similarity matching necessarily competing?
No, and we do not claim that they are. Please see the third paragraph of general response 1. A model optimizing both would conserve both quantities, so our data do not falsify similarity matching. Moreover, Fisher information depends only on the responses to neighbouring stimuli, so there exist transformations that conserve it while the representational similarity between distant stimuli changes. It is noteworthy that neither our simulations nor the data approach such solutions.

Q3. How robust are the empirical Fisher information estimates to covariance estimation and dimensionality reduction? Please see general response 2. Stability is unchanged across subspaces retaining 50 to 99% of residual variance (Figure A15). On the reliability benchmark, cross-day stability of Fisher information is closer to the within-session ceiling from split trial halves than to the across-animal floor (Figure 4C,D).

Q4. How does the proposed mechanism generalize beyond the simplified assumptions? Please see general response 3. On noise correlations in the learning dynamics specifically, we verified that the model with correlated output noise using the rule derived for general Σ in Appendix A.3 learns the Fisher information profile and its stability is preserved (Figure A1C).

Q5. What is the broader machine learning implication of the proposed learning rule?
Please see our response to W2 above.

AuthorsOpenReview ↗

Checking in on Our Rebuttal

Thank you again for your helpful feedback. We wanted to check if you have had a chance to look at our revisions, and if there are any remaining questions or clarifications we can provide before the discussion period closes.

I thank the authors for their substantial efforts in addressing the reviewers' concerns. The additional analyses, improved model comparisons, and more carefully calibrated claims have meaningfully strengthened the manuscript.

Regarding W1 and Q1–Q4, I acknowledge that the authors have addressed a substantial portion of my concerns. In particular, I appreciate the additional robustness analyses, the clarification of the relationship between efficient coding and similarity matching, and the explicit acknowledgment that the empirical observations establish consistency rather than mechanistic specificity. These revisions substantially improve the scientific rigor of the work.

However, my concern regarding W2 and Q5 appears to have been misunderstood.

I did not argue that computational neuroscience research is unsuitable for ICLR, nor did I request conventional machine learning benchmarks as a prerequisite for acceptance. I am well aware that ICLR welcomes theoretically motivated and neuroscience-inspired contributions.

My question was about the specific contribution of this work to machine learning, not whether neuroscience papers have historically appeared at ICLR. Citing historical precedents for the venue does not address this question.

I have also examined the newly added discussion in lines 533–539. While the proposed connection to continual learning is potentially interesting, it remains speculative. The authors suggest that preserving stimulus Fisher information without constraining individual parameters might reduce catastrophic forgetting, but neither theoretical analysis nor empirical evidence is provided to establish this property.

To clarify, I am asking the authors to identify what generalizable learning principle, theoretical result, or methodological insight this work establishes for machine learning beyond its neuroscientific interpretation. This need not involve additional benchmarks, but it does require a substantive argument grounded in the proposed formulation and its demonstrated properties.

In summary, I consider the concerns regarding the neuroscientific formulation and empirical validation substantially addressed. However, the concern regarding the paper's contribution to machine learning remains unresolved.

I appreciate the authors' revisions and remain open to reconsidering my assessment if they can substantively address this remaining concern.

AuthorsOpenReview ↗

Response to W2

We apologize for the misunderstanding and thank the reviewer clarifying their question and for the opportunity to revise our response to W2. The broader ML connection we offer is now in the Discussion, lines 496-504.

The learning rule proposed in equation (6) works because under isotropic noise, the expected squared temporal derivative of a representation, conditioned on the input variable that drives it, is proportional to the Fisher information about that variable. The noise variance enters only as the constant of proportionality, so for the deterministic networks typical of machine learning the same statistic is the Fisher information under any assumed isotropic noise level. Unsupervised algorithms like slow feature analysis and related slowness objectives penalize the squared temporal derivative on average, so learning invariance by suppressing temporal variation amounts to descending an information measure and sacrifices information about the variable driving the variation. The rule derived here instead performs gradient ascent on the same statistic in order to improve discriminability about that variable. As a corollary, in lagged or predictive formulations of slowness objectives, the timescale over which temporal variation is penalized would determine which variable the learned representation becomes invariant to. This insight could help inform the choice of the timescale in such machine learning algorithms.

We have removed the speculative discussion point about relation to continual learning due to space constraints. Please let us know if we can provide additional clarification.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank the reviewers for their thoughtful and constructive feedback. Below we address concerns raised by more than one reviewer and add fifteen supplementary analyses and three appendix sections to support our responses (W: Weakness, Q: Question).

  1. Stable representational similarity implies stable Fisher information, and the near-zero stability under similarity matching is an artifact (SXGs W1 and W2, qTBq W4, Vbib Q1 and W1)

Reviewer SXGs is correct. Under isotropic noise the Fisher information (J) at each stimulus is proportional to the squared distance between the responses to its neighbours, which is by definition, the dissimilarity between them, so conserved similarity implies a conserved J and the observation cannot separate the accounts. Related to this, qTBq’s concern that the near-zero correlation reflects two flat profiles is also correct. We rectify both by stating this equivalence in Appendix A.3 now and simulating that case heterogeneous input tuning (Figure A17).

But neurons in the brain typically exhibit correlated noise, and J(s) = μ′(s)ᵀΣ⁻¹μ′(s) would depend on the direction of the tuning slope relative to the noise, so preserving pairwise distances alone need not preserve J. This is why Fisher information in the recordings was estimated using each session’s measured covariance. We agree, however, that the original model simulations used isotropic noise and homogeneous input tuning, so the apparent dissociation merely reflected the noise around a flat fisher information profile. Therefore, we reran the comparison with correlated noise in both models, heterogeneous input tuning and at matched drift rates (Figure A16). Our original conclusions still hold (see Figure 4). Representational similarity is conserved in both models, whereas a stable Fisher information profile emerges under efficient coding but not under similarity matching. In this case the plasticity rule for efficient coding requires inverting the noise covariance which introduces nonlocal terms. We note this in the methods and discussion and leave a plausible biological implementation as an important future direction.

We agree with Vbib that the recordings establish consistency with efficient coding rather than showing that the brain achieves it, and we do not claim drift is produced only by efficient coding. Our finding is that the neural recordings in sensory regions conserve quantities beyond representational similarity and calls for expanding upon existing accounts. The additional conserved quantity is the one predicted by the proposed efficient coding objective. A model optimizing both would conserve both quantities, so our data do not distinguish the two. What they suggest is that similarity matching alone is insufficient, since it conserves similarity but not J.

2. Robustness of the Fisher information estimates to covariance estimation, and a scale-insensitive stability metric (Vbib Q3, qTBq W2)

Stability results are robust to the precise explained variance threshold and remain unchanged when the covariance is inverted within subspaces retaining 50 to 99% of residual variance (Figure A15). Because the Pearson correlation is insensitive to overall scale, we repeated the analysis using the normalized squared difference (Figure A21), and added within-session ceiling from split trial halves and across-animal floor for comparison (Figure 4C,D).

3. Generality beyond the simplifying assumptions (Vbib Q4, qTBq)

The fisher information allocation and its stability hold under a saturating output nonlinearity, with temporal rather than stimulus derivatives, with correlated noise in the learning dynamics (Figure A1), with Poisson-like scaling of response noise (Figure A4), with noisy inputs (Figure A7), across network sizes and prior sharpness (Figure A2), and for a range of network expansion ratios (Figure A9). Because the time derivative of a noisy response is not well defined at small step sizes, we also propose and simulate a variant of the rule that accumulates the pre- and postsynaptic derivative product in an eligibility trace, which converges to the optimal profile (Appendix A.5, Figure A3). Model comparison now also includes a non-negative variant of similarity matching (Figure A18).

4. Whether drift is driven by the learning rule or by the added noise, and the difference from earlier work (qTBq W5, Vbib Q4)

Drift also occurs without synaptic noise due the stochasticity of online learning, although more slowly (Figure A6), so it does not follow solely from the added noise. Drift is slower for frequent stimuli (Figure A5), as reported for familiar odorants in olfactory cortex. We agree that noise-driven diffusion on a solution manifold is not new, and the Discussion now states the difference from Pashakhanloo and Koulakov. So, the learning rule proposed here is a plausible explanation for the population metric that is conserved under drift rather than as a novel cause of drift itself.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.