acceptodds
Under review as a conference paper at ICLR 2027

Sparse Yet Adaptive: An Interpretable Transformer for Neural Population Readout

Abstract

How neural systems selectively read out reliable information from high-dimensional and redundant population activity remains a central problem in systems neuroscience. Conventional decoders can establish what information is present in neural activity, but provide limited insight into how that information may be flexibly read out. Here, we introduce a transformer-based framework that uses self-attention as an interpretable probe of population readout. Applied to large-scale macaque V1 recordings, the model revealed internal readout dynamics with clear biological correspondence: highly weighted neurons were tuned near the current stimulus, and removal of dominant readout neurons triggered recruitment of functionally similar substitutes, consistent with redundant and resilient population coding. We further applied the framework to sensory perturbations. Despite reduced contrast or moderate external noise, orientation information remained accurately recoverable, paralleling perceptual robustness. Attention analysis further suggested a possible mechanism for this stability: as sensory evidence degraded, accurate readout recruited additional neurons and weaker population interactions. Together, our framework moves neural decoding beyond black-box prediction toward an interpretable tool for uncovering computational strategies and generating mechanistic insights. Code is available at https://anonymous.4open.science/r/Transformer-model-for-accurate-orientation-coding-0751/.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Neural decoding, Self-attention, Macaque V1, Two-photon imaging

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_JzJQOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2(+1)
Read review
Paper summary

Premise of the work is that visual information is distributed in redundant high dimensional population responses in the brain. The goal of the paper is to extract useful principles of how can this information be read out (by downstream circuits) through the lens of a transformer-based decoder trained to reconstruct the sensory input from recorded neural responses. Specifically, its attentional weights are used as a proxy for what a readout circuit might do with those responses during brain function. This is applied to data from monkey V1, in two conditions (clean and noisy gabor gratings) yielding some features that the authors consider realistic. Nonetheless, the very dynamic nature of the extracted readout makes this interpretation suspect from the perspective of biology.

Critical strengths

Interesting data in terms of effects of pixel level noise on V1 encoding at the population level

Critical weaknesses

A) The idea that the weights reflect a biological readout is to me fundamentally suspect, even if (as the authors advise) interpreted in terms of functional rather than synaptic connectivity.

The robustness of the representation coming from a distributed code means that potentially multiple readouts could get the job done. The argument here is that the self-attention can more directly take into account the coherence pattern of the population responses to decide what's a minimum useful subset of neurons that tell me what I need to know. I see at least two fundamental conceptual issues with this interpretation: 1) the computation of the self attention mask takes into account the full population, data that a downstream circuit is not expected to have. 2) related, but the readout being input specific is equally problematic since a downstream circuit cannot decide which neurons to read from until knowing what stimulus they are looking at, which fundamentally defeats the purpose of downstream information transmission. In other words, the downstream circuit cannot have enough knowledge to know which neurons to read from. As a separate issue, since there are multiple readouts that work, the degeneracy will likely be resolved via the inductive biases of the model, which are in no way set to reflect brain properties. It's not clear to me why any alignment between solutions would be more than happenstance.

B) Numerical comparisons to strawman models is superficial. Not clear how if any effort was put into making alternative nonlinear decoders work.

C) not clear why the results are deemed biologically sensible, or how much of the pattern preservation can be explained from the functionally-based way the neurons were indexed in the first place. Similarly -- redundancy is part of the premise so the fact that that is found as part of a more complex analysis seems like busywork.

D) found the figures generally unintuitive and difficult to see and interpret

E) other transformer based encoder-decoder solutions exist (e.g recent papers from Tolias lab), pixel level decoding has been done for retinal responses in a number of ways (e.g. Chichilnisky, Marre), the novelty of the approach is overstated

Questions for authors

What is the ground truth I, exactly? Regarding image reconstruction (as it appears on the screen?): Is the field of view large enough that the receptive field of the cells recorded span the full screen? how is that verified? are the gabors full field? What is the spatial resolution for the images to reconstruct?

What is the relationship between df/f in the data and the variable r?

Are dynamics over time something that the model is sensitive to?

How is the orientation selectivity of the cells determined (for their ordering in the model)?

Model controls: why have a nonlinear decoder of orientation as a label but not an MLP version of pixel level reconstruction?

The SSIM metric is not clearly providing a sense of the quality of the reconstruction, would it be possible to see some example reconstructed images for direct comparison? Also I can believe that a linear decoder out of the box does worse in terms of orientation estimation but why is the MLP version not any better? I find the control numbers counterintuitive. please explain

Missing definitions for variables, e.g. HxW is presumably the height and width of the image in pixels?

wouldn;t the alignment of high weights with neural tuning trivially be true even for a linear decoder with a weak sparsity constraint that picks different neurons across seeds? why is this finding interesting?

At which point in the analysis is the alternative linear decoder of I used for comparison? I seem to have missed that in the text

3 comments

Concern

A: We thank the reviewer for these important points.

(1–2) Access to the full population. Attention weights are computed from the population response itself, so no knowledge of stimulus identity is required. However, we agree that this computation requires access to the full population, which is unavailable to downstream circuits. We therefore do not propose attention as a biological mechanism. "Sparsity" refers to the effective readout weighting. Our claim is limited to showing that, under redundant coding, a selective readout exists that is efficient and stable. We now state this explicitly in the Limitations.

(3) Degeneracy and inductive biases. We agree that many readouts achieve similar accuracy and that the model's inductive biases select among them. Indeed, shared embeddings reach comparable accuracy without tuning alignment (Fig. 3A), so accuracy alone does not imply biological correspondence. The alignment we report, however, is unlikely to be happenstance. It is measured against independently estimated tuning never provided to the model, persists after shuffling neuron order (Fig. 3D), and is consistent across FOVs.

We interpret it as a property of the solution favored by this architecture, not a uniquely biological one, and have revised the Discussion to frame these results as a testable hypothesis about selective readout requiring causal, circuit-level validation.

B: We selected established linear, MLP, and no-attention baselines, clarified their rationale (Section 3.4), and detailed optimization (Appendix A.3).

C: Functional tuning is known properties of V1. Our goal is not to rediscover them, but to ask how a readout can selectively use redundantly coded information while remaining efficient and stable. Biological sensibility. We assess it through stimulus–tuning alignment: the model's selected neurons match independently measured orientation tuning never given to the model. This is consistent with causal evidence that functionally defined ensembles can drive perception (Carrillo-Reid et al., 2019; Marshel et al., 2019), but we do not claim it establishes a biological mechanism. Neuron indexing. To test whether the pattern follows from the functional ordering, we randomly shuffled neuron positions and retrained the model. Stimulus–tuning alignment and sparse selection were largely preserved (Fig. 3D), so are not imposed by the original ordering. Redundancy. Redundancy is a premise, not a finding. What we add is how the readout behaves under it: after removing the highest-weighted neurons and retraining, functionally similar neurons took their place and decoding accuracy was maintained (Fig. 3B–C). This shows that the sparse selective readout is stable under redundant coding.

We frame these results as a testable computational account, not an established biological mechanism.

E: Our main contribution is using stimulus-dependent attention maps, which conventional decoders do not explicitly provide, to investigate computational strategies underlying neural population readout, rather than proposing a novel Transformer architecture or image-reconstruction method.

Q1: The target was a 79 × 79-pixel (~1°) Gabor within each FOV's mapped pRF, not full-screen (Appendix A.1, Fig. A.2A).

Q2: r∈RN\mathbf{r} \in \mathbb{R}^{N} is the vector of ΔF/F\Delta F/F responses of all NN neurons (Sec. 2.2, lines 133–138).

Q3: The model does not explicitly capture temporal dynamics. Fluorescence signals were averaged over the baseline (−500 to 0 ms) and response (500–1,000 ms) windows to calculate a single ΔF/F value per neuron per trial.

Q4: Neurons were ordered by preferred orientation estimated from training-only Gaussian fits, and this ordering remained fixed across all conditions. Details are provided in Section 3.2 and Appendix A.2.

Q5: In fact, the no-attention baseline included in our study is an MLP-based image reconstruction model. We have clarified this baseline in the revised Section 3.4 (Baselines) and provided further architectural details in the Appendix A.3 (lines 789-795).

Q6: Example stimuli and reconstructions are now shown in Fig. A.2B. Although MLPs have greater nonlinear capacity, this need not improve generalization, likely because our limited dataset cannot exploit it. To further investigate this, MLPs with 1–5 hidden layers gave median orientation errors of 2.922°, 3.581°, 3.807°, 3.162°, and 4.475°, versus 2.83° for the linear decoder, so added depth did not improve decoding on our dataset.

Q7: Yes, H=W=79H = W = 79.

Q8: Linear decoder weights showed no clear systematic relationship with orientation preference (Appendix B.5). Stimulus–tuning alignment is a criterion for biological relevance, not a novel finding itself (see response to L92v, Concern 2).

Q9: The alternative image-reconstruction decoder is No-attention baseline. We have clarified this in Section 3.4 and provided implementation details in Appendix A.3.

Reviewer_JzJQOpenReview ↗

Rebuttal received

I thank the authors for the clarifications but I still have a fundamental problem with the formalism and what it can possibly teach us about the nature of neural coding. I will maintain my score.

We thank the reviewer for their comments. We would like to clarify two aspects of our contribution.

First, regarding the computational formalism, our architecture was designed to mathematically characterize two key properties of neural populations: neuron-specific embeddings represent individual neuronal identities and functional heterogeneity, while self-attention captures response-dependent interactions among neuronal representations. Different population responses naturally produce different attention maps through neuron-specific embedding and self-attention, allowing the model to automatically assign neuronal weights without prior knowledge of stimulus identity or which neurons should be selected. Unlike conventional decoders with fixed (stimulus-independent) readout weights, our framework provides explicit stimulus-dependent attention maps, allowing us to examine how neuronal contributions and interactions vary across stimuli. This advances the interpretability of population readout beyond conventional decoding analyses. Previous computational studies have also explored potential biological implementations of Transformer-like computations through neuron–astrocyte networks (Kozachkov et al., 2023).

Second, regarding what our findings teach us about neural coding, our results highlight a computational principle of redundant encoding coupled with sparse effective readout, complementing classical efficient- and sparse-coding perspectives (Barlow, 1972). Building on the broader ideas of selective information gating and flexible routing in neural circuits (Yang et al., 2016; Wang & Yang, 2018), our study takes a further step by quantitatively characterizing stimulus-dependent selective readout using real V1 population responses and examining its flexibility and robustness through in silico neuronal removal and model retraining. Specifically, our results further show that distributed population information can support selective, stimulus-dependent readout, while redundancy allows functionally similar neurons to support alternative readout solutions following neuronal removal and model retraining, thereby preserving decoding performance.

Thus, our study advances both methodology and computational understanding of neural population readout. Methodologically, it enables interpretable analysis of stimulus-dependent neuronal weighting beyond decoding accuracy. Conceptually, it illustrates how redundant encoding can support sparse, flexible, and robust effective readout. Our findings provide a data-grounded, testable computational hypothesis for future causal investigation.

We sincerely thank the reviewer again for their thoughtful comments on both the conceptual framework and methodological details, particularly the experimental procedures, which have helped us clarify our arguments and improve the rigor and presentation of our manuscript.

References

  1. Barlow, H. B. (1972). Single units and sensation: A neuron doctrine for perceptual psychology? Perception, 1(4), 371–394. https://doi.org/10.1068/p010371

  2. Kozachkov, L., Kastanenka, K. V., & Krotov, D. (2023). Building transformers from neurons and astrocytes. Proceedings of the National Academy of Sciences, 120(34), e2219150120. https://doi.org/10.1073/pnas.2219150120

  3. Yang, G. R., Murray, J. D., & Wang, X.-J. (2016). A dendritic disinhibitory circuit mechanism for pathway-specific gating. Nature Communications, 7, 12815. https://doi.org/10.1038/ncomms12815

  4. Wang, X.-J., & Yang, G. R. (2018). A disinhibitory circuit motif and flexible information routing in the brain. Current Opinion in Neurobiology, 49, 75–83. https://doi.org/10.1016/j.conb.2018.01.002

Reviewer_Qe1bOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper applies a transformer architecture with self-attention to examine information content among populations of visual cortex neurons recorded with 2 photon calcium imaging while macaques view oriented gratings under different levels of stimulus contrast and signal-to-noise ratio. The transformer architecture is compared to some ‘baseline’ decoders, and seems to result in very accurate stimulus reconstructions (as quantified by the orientation content reconstructed). The paper additionally includes analyses testing how self-attention jointly selects neurons for readout for each type of stimulus, the impact of joint vs stimulus-specific embedding spaces, and how model choices impact the relative tuning properties of neurons selected for readout. The work overall goes beyond current approaches, and is particularly innovative in its use of animal imaging data, which is a major strength beyond the common use of noninvasive measurements like fMRI, EEG, MEG from human participants.

Critical strengths
  1. The dataset is extremely interesting, encompassing hundreds of simultaneously-recorded neurons as measured from macaque visual cortex with 2-photon calcium imaging methods. This affords a level of population coverage and precision that’s rarely achieved, and makes these results particularly impactful. The stimulus set chosen is also especially useful and allows for close links to previous work on visual encoding of simplified visual features
  2. Decoding performance (quantified by median orientation decoding error) with the transformer architecture is exceptional (<1 deg error)
  3. The paper begins with a compelling argument that readout of the ‘full’ stimulus, rather than a reduced feature dimension among many, may be a strong target for understanding neural information codes. This is not commonly addressed, and this work makes a substantial contribution towards this goal beyond studies that focus only on an encoded feature dimension (e.g., orientation, rather than image of a Gabor).
Critical weaknesses
  1. While much of the approach employed seems sensible, there are some choices that are not clearly described or justified. Primarily, the lack of description for how neurons are sorted according to their feature preference (Cross-validated, such that preference is sorted based on training data only? Using separate tuning function estimates?) in the Positional Encoding step (Fig. 1; Eq 2) is a critical missing detail. Ensuring independence of any data used for fitting the transformer (including data used to sort neurons) from the data used for image reconstruction/decoding is essential.
Questions for authors
  1. Why were data from the lowest-contrast condition not further analyzed? (pg 2, line 105)
  2. The paper reports primarily median orientation decoding errors, but the shape of the error distribution can also be informative. What do decoding error distributions and/or decoding error scatterplots (decoded orientation vs presented orientation) look like, and do they have any informative structure?
  3. The paper does a thorough job considering many model forms, including a comparison of stimulus-specific embedding estimates for each neuron to a shared embedding across stimuli for each neuron (Fig. 3). While this may be a standard approach in designing AI models, how would the stimulus-specific embedding work in practice in a neural system? Is there previous literature/modeling work the authors could cite (or are there other arguments the authors could make) supporting this structural choice?
  4. For image-based decoding, the authors use a spatial gradient measure applied to BOTH the reconstructed image and the decoded image to compare the decoder accuracy (pg 14, Lines 716-17) . Why is the target image orientation extracted from the gradient method, rather than with the ‘true’ stimulus orientation presented to the animal? How distributed are the gradient-based quantifications of stimulus orientation compared to the true orientation embedded within the stimulus, and does this vary as a function of stimulus orientation? Do the results presented throughout hold when ‘true’ stimulus orientation is used for computing decoding accuracy, rather than gradient-based stimulus orientation estimates?

Suggestion:

  1. For Fig. 4A, I would suggest the authors replace the categorical colors of each cell of each image with a continuous colorscale reflecting that cell’s decoding accuracy. The distinction between cells, since continuous, could be illustrated with a dark dividing line, or asterisks, or some other visualization strategy. The data as presented presently imply a sharp distinction between those values when they are really continuous.
2 comments

Concern 1: We confirm that neuronal orientation preferences were estimated exclusively from the training data, without using any held-out test data. For each neuron, we estimated its preferred orientation by fitting a Gaussian tuning curve using training-set responses and defined the fitted peak as its preferred orientation. Neurons were then sorted accordingly. Once established, this neuronal ordering remained fixed across all stimulus conditions and throughout both model training and testing. We have clarified the estimation and fixed neuronal ordering procedures in Section 3.2 (Positional encoding) and provided further details of the tuning-curve fitting in Appendix A.2.

Q1: The 3.125% contrast condition was initially excluded because data were available for only one of the three FOVs. We have now included the available data in the revised Fig. 4. Importantly, including this condition does not change the overall trends or conclusions.

Q2: We have added decoding error distributions and scatterplots of decoded versus presented orientations to the revised Figure B.2. The transformer model exhibits a right-skewed error distribution, with most errors concentrated near zero and only a small fraction extending into a longer tail. Compared with the baseline models, it also shows a narrower error distribution and smaller maximum errors, indicating more consistent decoding performance across stimuli. We have incorporated these findings into the revised Results section (lines 276–278).

Q3: First, a clarification of terminology: our design uses neuron-specific embeddings that are fixed across stimulus conditions, not stimulus-specific embeddings. What varies with the stimulus is each neuron’s response vector (and the resulting stimulus-dependent self-attention map); the embedding vector itself does not change across conditions.

This design is motivated by the functional heterogeneity of V1 neurons, which exhibit distinct tuning preferences and receptive-field properties (Ringach et al., 2002). Using a fully shared embedding may limit the model's ability to explicitly capture these neuron-specific functional differences. In contrast, neuron-specific embeddings preserve individual neuronal identities, while shared attention computations enable population-level interactions.

Importantly, similar neuron-specific parameterizations have been widely adopted in biologically motivated neural models. Klindt et al. (2017) introduced the “what–where” factorized readout, combining shared feature extraction with neuron-specific spatial and feature weights. Burg et al. (2021) applied a similar framework to macaque V1 responses. More recently, Wang et al. (2025) demonstrated that neuron-specific readout weights exhibit stable functional characteristics across recording sessions and contain information about neuronal anatomical properties. Furthermore, Willeke et al. (2026), from Tolias and colleagues, introduced OmniMouse, a multimodal Transformer model of mouse visual cortex that explicitly incorporates learned neuron-identity embeddings alongside session- and animal-specific embeddings. These findings provide biological support for using neuron-specific representations to capture stable functional differences among neurons.

Our embeddings are computational representations of neuronal identity and functional heterogeneity, rather than literal implementations of biological mechanisms. Our shared-embedding control further supports this design choice: although the shared-embedding model achieved comparable decoding accuracy, it failed to exhibit systematic alignment between highly weighted neurons and the presented stimulus orientation (Fig. 3A). This indicates that accurate decoding alone does not guarantee biologically meaningful stimulus–tuning correspondence, highlighting the value of neuron-specific representations for interpretable population readout.

We have clarified the motivation for neuron-specific embeddings and incorporated the relevant references in the revised manuscript (Section 3.2, lines 174–178).

Q4: We confirmed that orientations extracted from the original Gabor images using the gradient-based method closely matched the ground-truth stimulus orientations. Since the orientations of reconstructed images must be estimated from the images themselves, we initially applied the same gradient-based method to both the original and reconstructed images to ensure methodological consistency. To directly address the reviewer's concern, we also recalculated decoding errors using the ground-truth stimulus orientations as references. The median decoding errors were 0.12° and 0.15° when using gradient-derived and ground-truth orientations as references, respectively. These results confirm that the choice of reference has only a minor effect on decoding accuracy and does not change our main conclusions.

Suggestion 1: We have revised Fig. 4A.

I appreciate the authors' careful attention to my previous comments. The paper is improved, but in my view, still sits within the "3" category, so I will keep my score.

Reviewer_L92vOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces a transformer-based framework that uses self-attention as an interpretable probe of population readout. Applied to large-scale macaque V1 recordings, the paper reports that highly weighted neurons tend to prefer orientations close to the presented stimulus and that a small subset of neurons can preserve accurate orientation reconstruction. After removing dominant neurons and retraining the model, alternative neurons can support accurate reconstruction. The paper may provide an interpretable tool for revealing biological visual encoding strategies

Critical strengths
  1. The scientific question the paper tries to address is quite important and interesting. Going beyond the conventional black-box decoding method, an interpretable framework that uses a stimulus-dependent attention mechanism is potentially useful approach for revealing how information may be selectively read out from high-dimensional neural populations and the underlying computation principle.
  2. The paper investigates biological visual representations, which is quite intereseting that go beyond standard decoding benchmarks, particularly the coexistence of sparsity and redundancy and the recruitment of additional neurons as sensory signal quality deteriorates. These ideas are consistent with a growing body of recent computational and experimental work. Since the study relies on a private dataset, it provides valuable experimental evidence that could further support and advance our understanding of this field.
Critical weaknesses

Several issues substantially undermine the scientific rigor and publishability of this work. I have combined my weaknesses and questions below. I am very interested in neuroscience, and would be glad to see high-quality neuroscience research published at ICLR.

  1. In Sec. 3.2, the author claims Neurons are ordered according to their preferred orientations before a fixed sinusoidal positional encoding is added. Thus, orientation preference information is already used by the model, introducing a potential architectural bias. The observation that neurons with high attention scores prefer orientations close to the current stimulus therefore does not establish the authors' claim.
  2. Since the primary task is orientation decoding, assigning higher attention scores to neurons tuned to the stimulus orientation is unsurprising. More importantly, the computation of attention scores has already used the activities of the entire neural population. Therefore, even if only 2–3 neurons receive high attention scores, this does not imply that the decoder relies exclusively on information from those neurons.
  3. Removing highly attended neurons and retraining the model does not demonstrate the claimed adaptive coding mechanism. Given the simplicity of the task and the limited stimulus space, the remaining neurons may still provide sufficient information to identify the stimulus, potentially even through exclusion. This result demonstrates decoding robustness rather than adaptive neural coding. Stronger evidence for redundant population coding would require more complex stimuli, such as natural images, where neural responses may not uniquely determine the stimulus.
Questions for authors

see weakness.

1 comment

We thank the reviewer for their constructive comments, which motivated an additional neuronal-order shuffling control and helped us clarify the distinction between model-level readout reorganization and biological adaptation. Corresponding revisions are highlighted in blue.

Concern 1:

To determine whether orientation-based neuronal ordering influenced the observed stimulus–tuning alignment, we performed an additional control experiment in which neuronal positions within the input sequence were randomly shuffled and the model was retrained. Decoding accuracy remained high (median absolute orientation error = 0.11°) and these neurons continued to exhibit orientation preferences close to the presented stimuli, with median differences of approximately 5°–10° (Fig. 3D). These results demonstrate that sparse, stimulus-aligned selective readout does not depend on the orientation-based ordering prior. We have also added an explanation of the rationale for orientation-based neuronal ordering. This design was introduced as a biologically motivated inductive bias, inspired by the functional clustering of orientation preferences in macaque V1. This arrangement allows neurons with similar orientation preferences to receive similar positional encodings.

We have expanded Section 3.2 (Positional encoding) to clarify the rationale for neuronal ordering and provide further details of the ordering procedure. We have also added Appendix A.2 to describe how neuronal orientation preferences were estimated and Section 4.5 and Fig. 3D to present the neuronal-order shuffling control and its results.

Concern 2:

The correspondence between attention weights and neuronal orientation preferences provides a biologically meaningful criterion for evaluating the model-derived readout strategy. Previous causal studies have demonstrated that targeted activation of functionally defined neuronal ensembles in visual cortex can influence visual perception and behavior (Carrillo-Reid et al., 2019; Marshel et al., 2019). Importantly, such alignment is not guaranteed by accurate decoding: the shared-embedding model achieved comparable accuracy but failed to exhibit systematic stimulus–tuning alignment (Fig. 3A). Unlike conventional linear decoders with fixed readout weights, our model explicitly captures stimulus-dependent, nonuniform weighting of neuronal signals, providing an interpretable readout strategy.

Second, self-attention is computed from the entire neuronal population, and the model is trained with all recorded neurons. However, full-population input does not imply uniform readout. Instead, the model selectively emphasizes a sparse subset of task-relevant neuronal signals. To quantify which signals the trained model preferentially relies on and how many are needed to achieve a given accuracy, we progressively retained highly weighted neurons and their incoming attention-defined interactions while evaluating reconstruction performance (Appendix Fig. B.3). Retaining approximately 2–3 highly weighted neurons per stimulus and their incoming interactions was sufficient to achieve an orientation error below 5° (Fig. 2D). This does not imply that these neurons independently support decoding, as they receive information from other neurons through attention. Including these inward connected neurons, the effective readout involved 22.7 ± 5.31% of the population, with effective interactions accounting for only 0.021 ± 0.003% of all possible pairwise interactions. We have added this quantitative clarification to the revised manuscript (lines 309-311).

Thus, our central finding is that information distributed across a redundant neuronal population can be selectively read out through a sparse, stimulus-dependent weighting structure, rather than requiring all neurons to contribute equally. This provides an interpretable computational account of how sparse selective readout can coexist with distributed redundant population encoding.

Concern 3:

This is an important point. The term "adaptive" could be misleading. Our results demonstrate model-based decoding robustness supported by redundant population information. We have therefore revised the manuscript to clarify this distinction and avoid interpreting model retraining as a biological adaptive mechanism. We also agree that more complex stimuli, such as natural images, would provide a more stringent test of the generalizability of this strategy. Nevertheless, orientation selectivity is a canonical and well-characterized property of V1, making it a biologically grounded and controlled setting for investigating how distributed population information can be selectively read out. We have included this limitation and the extension to more complex visual stimuli in the revised Discussion (lines 521-524).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.