acceptodds
Under review as a conference paper at ICLR 2027

RetinaPlex: A Large-Scale Biologically Constrained Retinal Model for Parallel Visual Coding

Abstract

Visual perception depends on sensory populations that represent ongoing scenes while retaining information about recent changes. How retinal diversity organizes these complementary signals for downstream computation remains incompletely understood. Here we introduce RetinaPlex, a biologically constrained retinal model comprising approximately ten million modeled retinal units and supporting 4K visual input. The model combines spatial organization constrained by human photoreceptor statistics and informed by primate retinal connectivity with physiologically inspired temporal dynamics. Across controlled stimuli and natural-video experiments, sustained and transient bipolar populations support low-capacity readout of current scene state and recent visual history from a single population snapshot, with retinal parameters held fixed. Distinct histories remain distinguishable after input sequences converge to identical continuing visual input. Model interventions show that retained-response differences shape early state–history geometry, whereas adaptation timescales modulate the relative contribution of history across pathways over time. Reanalysis of public primate retinal recordings reveals a concordant association between independently measured temporal response phenotypes and relative current-versus-recent coding bias. Comparisons with frame-like and event-like interfaces further characterize access to both information sources. Together, these findings establish RetinaPlex as a platform for investigating retinal population coding and suggest biologically grounded principles for sensory representations that jointly preserve scene state and recent change in NeuroAI and neuromorphic vision.

Primary area:
applications to neuroscience & cognitive science
Keywords:
biological neural networks, sensory coding, sensory neuroscience, computational neuroscience, retina, NeuroAI

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_Mv64OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose RetinaPlex, which is a biologically (e.g. semi-connectome) constrained model of the primate retina that operates on 4K input. They then show that, concordant with marmoset retinal recordings, that in their model, retained response differences in their model separate current and past signals, and adaptation in the model redistributes past signals across pathways.

Critical strengths

I think more models of primate retina are needed, and so this to me is a considerable strength of the work, as prior high-fidelity retinal models have been in other animal models (e.g. salamander as in Maheswaranathan et al. 2023)

Critical weaknesses
  • It is not clear to me what mechanisms this work better predicts that wasn't already present in existing data (e.g. Kruppel et al. 2023). In other words, it is presently unclear what the benefit you get from using this model that this better enables for science (see my questions below in the next section).

  • It would be important to compare to other retinal models (e.g. Maheswaranathan et al., 2023 and OpenRetina) to see if their model better predicts the phenomena in Figs 3-4, especially in marmoset retinal data that they compare to of Kruppel et al. 2023 in Fig 4c.

Questions for authors
  • Can you clarify how response difference between current and past signals is novel in the retina? After all, we know that there is considerable temporal filtering in the retina, so this seems to already be well-known

  • Likewise, can you clarify how the adaptation in the model isn't already explained by known adaptation mechanisms, like contrast adaptation, luminance and mean adaptation, spatial adaptation, etc.?

  • Similarly, can your model recapitulate other known mechanisms in the retina, as well as predict anything new, like Tanaka et al. NeurIPS 2019 (https://pubmed.ncbi.nlm.nih.gov/35283616/) did with the omitted stimulus response (OSR)?

1 comment
AuthorsOpenReview ↗

Authors' Response: Mechanistic Scope, Model Comparisons, and Falsifiable Predictions

Response to Reviewer Mv64:

We greatly value your expertise and high-confidence assessment. Your question—what does this model enable that existing data or models do not?—is exactly the right one.

M1. Is the current/history response difference new? What does the model enable scientifically?

Response. We agree that retina-dependent temporal filtering and history-sensitive RGC responses precede this work. We do not claim to have discovered history dependence itself. The tested question is how a spatially structured population makes present and recent variables simultaneously accessible, and which changes to effective pathway dynamics alter this geometry.

Unlike unconstrained observations, our fork–rejoin design controls the subsequent input, and interventions isolate retention and adaptation. To further demonstrate the scientific benefit, we added Section 4.6, showing that a frozen, two-channel S/T state serves as a computationally viable preprocessing module for a real-video downstream task. This confirms that the S/T organization is not merely a generic filter, but a highly structured, low-bandwidth representation capable of supporting actual decisions.

M2. How does RetinaPlex compare with Maheswaranathan et al. and OpenRetina?

Response. We have now performed a shared-stimulus representation-phenotype comparison. Phenotype groups are determined using independent step responses before the current/preceding decoding task (100–150 ms).

For RetinaPlex, S and T groups achieve 100% current decoding, with 46.48% vs. 94.53% history decoding. Equalizing rr changes history accuracies to 82.81% vs. 60.55%. A generic four-state control also shows a related effect, while OpenRetina core and Raw Two-IIR reach 100% on all four components, creating a ceiling effect.

However, reaching 100% does not mean black-box models provide the same insight. While OpenRetina achieves high decoding, our model disentangles the S/T contributions, providing mechanistic insight that black-box models cannot offer. Furthermore, as shown in Sec. 4.6, RetinaPlex compresses 19 frames into just two biologically interpretable maps, matching high-dimensional models. This suggests complex temporal filtering is highly organized and can be efficiently captured at the early bipolar-cell stage. We also analyzed native GC outputs; they are not conflated with hidden C2/core features.

M3. Is the modeled “adaptation” already explained by known mechanisms?

Response. The adaptation state in Eq. (4) is an effective temporal variable. It is not identified with a specific molecular process, presynaptic resource-depletion, or comprehensive luminance adaptation circuit.

Our controlled light-history analysis shows that under a strictly common drive, the measured history-by-suffix-luminance interaction is numerically zero, as expected analytically. For learned representations, a nonzero interaction appears, but a fixed, memoryless softplus also creates nonzero interaction. Thus, that statistic alone cannot identify biological feedback. We restrict our physiological claim to effective dynamics and their population-coding consequences, not a newly established form of retinal adaptation.

M4. Can the model make a genuinely new, independently validated biological prediction, e.g., an OSR?

Response. We agree that predicting an OSR-like phenomenon would constitute stronger cellular validation. We have not tested OSR and do not claim to have predicted it.

Our tests establish a different, falsifiable result: for prespecified manipulations, changing retention and adaptation produces distinct consequences for readout geometry. A neuron-level prospective validation remains a separate target. However, our fork–rejoin framework provides a generative tool to propose such hypotheses in future work. The present paper's claims are strictly about population representation, intervention, and conditional downstream utility.

Summary on Relation to Krüppel et al. (2023): Finally, regarding the distinction between our model and existing empirical data: while Krüppel et al. established the biological phenomenon of heterogeneous history sensitivity, our model provides a causal, generative explanation for how this heterogeneity emerges from specific S/T pathway dynamics.

We emphasize that the anatomical priors act as an inductive bias to reveal computational principles, rather than a recovered biological ground truth. This aligns with the classic computational neuroscience approach: using constrained models to propose testable mechanistic hypotheses (such as the distinct roles of retention and adaptation) that are difficult to isolate purely through experimental observation. We hope this clarifies the complementary relationship between our theoretical platform and empirical recordings.

Reviewer_6UYdOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The authors introduce RetinaPlex, a large scale biologically constrained retinal model for parallel visual coding. RetinaPlex is a model that is trained to convert high res images (4K) to real biological ganglion cell recordings. Their motivating question is a deeper understanding of how the retinal population encodes persistent information in the ongoing scene at the same time as encoding recent updates to the scene. Their main contributions are stated as the contribution of the RetinaPlex model, the fact that the retinal population can be used to readout information about ongoing scenes as well as recent updates, and a mechanistic understanding of how the sustained and transient populations can be used to distinguish between degenerate potential past states for the current state.

Critical strengths
  • scale of the model is well sized for a retinal model
  • interpretability on models trained on biological data is a great direction for neuroscience on the whole
Critical weaknesses

My main concern with this paper is that the motivating question, model contributions, and findings are all catered to a neuroscience audience. Where the paper could have made contributions would be in showing that retinaplex improves visual processing accuracy or efficiency for multi-modal models, unlocks some new understanding of state-of-the-art vision/multi-modal models, or improves existing brain-computer interfacing (BCI) retinal implants in some way. As it is, the main motivation for the paper is retina-centric, which the vast majority of people at neurips do not care about outside of its potential significance for something like what I just mentioned. Thus, I cannot give a score of acceptance with the current framing and results.

What could change my mind: the framing would need to demonstrate relevance to AI improvements or BCI improvements in some way. And the supporting experiments would obviously need to show this as well. The problem is that this would likely require an entirely new paper which is against the spirit of the review process.

Specific (more minor) weaknesses:

  • unclear significance to the field of AI and ML
  • prohibitively dense jargon
  • lack of methodological clarity:
    • did you train on primate retinal recordings? Is this a population-level model or an individual-level model? Did the bipolar and ganglion cell recordings come from the same animal?
  • Figure 2c doesn't support claims: I would say the readout accuracy is low enough to debate the claim that it "remains readable early after rejoin." pg 7 line 350.
  • failed to introduce S/T before using in introduction
Questions for authors
  • does retinaplex, or the paper more generally, help AI or BCI in some way?
  • was retinaplex trained on primate retinal recordings?
  • what type of recording device was used for the the cell recordings?
  • did all recordings used for a single model come from the same organism?
2 comments
AuthorsOpenReview ↗

Authors' Response: Addressing AI Relevance, Downstream Utility, and Methodological Clarifications

Response to Reviewer 6UYd:

We thank you for your direct feedback regarding AI relevance and clarity. We realize the original submission did not adequately demonstrate downstream utility, and we have taken concrete steps to fix this.

U1. How does RetinaPlex help AI/ML, rather than only retinal neuroscience?

Response. We added §4.6, Figure 5 and Table 2 (pp. 9–10). Using real 25-Hz traffic video, the new task predicts whether an annotated vehicle bounding box first intersects a fixed circular zone within the next 200 ms. A frozen retinal front end processes 19 frames and exposes only two current spatial maps (S and T) to a trained ResNet18-GN decoder. Native-cell S/T achieves a source-macro AUROC of 0.729 ± 0.035, compared with 0.648 ± 0.050 for a current-frame baseline; it wins in 9 of the 10 paired seeds.

Crucially, in response to the suggestion for stronger baselines, we compared against the bipolar-like intermediate layers of Maheswaranathan et al. (0.717) and OpenRetina core (0.777). This highlights a key architectural insight: the efficient organization of present and recent information is already established at the early, bipolar-cell stage. RetinaPlex achieves this competitive performance using only 2 channels, compared to OpenRetina's 64 channels. This proves that a low-bandwidth state established early in the visual hierarchy can support complex temporal reasoning without requiring downstream complexity.

U2. What is the AI interface advantage of only two S/T channels, and is efficiency actually demonstrated?

Response. The added downstream experiment uses two current 128×128 S/T response maps, generated from the preceding video by fixed spatial connections. The decoder does not receive previous S/T maps or video frames. Under the same 128×128 grid and datatype, these two maps contain 9.5 times fewer scalar interface values than a 19-frame stack. This is a precise statement about the decoder input footprint; it is not a measurement of complete system runtime or energy cost.

The two channels are scientifically interpretable because they are defined by separate sustained/transient population dynamics rather than chosen as an arbitrary learned bottleneck. The revised §3.3 and Figure 5 distinguish this task interface from the 128D analytical snapshot, and Table 2 states the channel counts for all methods.

U3. Was the model trained on primate recordings? Were BC/GC recordings paired?

Response. No primate spike recordings were used to train the fixed RetinaPlex front end. Its anatomical priors and effective temporal parameters were specified independently. Low-capacity readouts are fitted on development data; the new ResNet18-GN decoder is trained separately using real-video contact labels. The marmoset recordings are an independent reanalysis of published neural data, not training targets. RetinaPlex is a population-level simulation.

The biological analysis covers four ex vivo retinal recordings from four marmosets using MultiChannel Systems planar multielectrode arrays (60 or 252 electrodes). There are no paired BC and GC recordings from the same animal in this study, and model BC units are not claimed to be registered to those real RGCs. We have clarified this provenance in the revised Methods.

U4. Does Figure 2c actually support the claim that recent information remains readable early after rejoin?

Response. We agree that “readable” must be quantitatively qualified. At the prespecified 125-ms time point, frozen readouts obtain 87.5% current-state accuracy (35/40) and 80.0% history accuracy (32/40) on five held-out video sources. Individual-source history accuracies are 75.0%, 75.0%, 75.0%, 87.5%, and 87.5%. A source-level leave-one-out check puts the history average between 78.125% and 81.25%. These values demonstrate above-chance, finite-time accessibility but not flawless reconstruction. History accuracy drops to 65.0% at 208.3 ms. We have revised §4.1 to state the source range and finite-time scope.

U5. Presentation and terminology.

Response. We have revised the first use to sustained (S) and transient (T), distinguished the full platform from the cone-driven S/T core evaluated in this paper, and distinguished the 128-dimensional analytical snapshot from the two spatial task maps. Figure 5 makes the input–front-end–decoder separation explicit. The new 128×128 shared-information task is not a direct test of native-4K accuracy or end-to-end speed; these clarifications address the notation ambiguities without changing the core architecture.

The authors have addressed many of my critiques, so I'm raising my score.

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

The paper asks whether activity at a single time point in a simulated retina can identify both the current input condition and the order of preceding stimuli. RetinaPlex receives native 4K luminance video and contains approximately ten million photoreceptor and bipolar spatial units. Its cell distributions and connections are informed by human and primate studies, and predefined equations describe how responses change over time. The main experiments use cone-driven sustained/transient (S/T) populations. Each branch is summarized over an 8-by-8 spatial grid, giving a 128-dimensional current population snapshot for downstream classification.

The principal classification experiments superimpose artificial spatial patterns on public DAVIS videos. Separate binary readouts identify the current contrast-modulation condition and whether earlier patterns occurred in AB or BA order. Paired histories receive identical subsequent video input. Further comparisons such as ABC/BAC also match the event set and final event, changing only the earlier order. These experiments directly test decoding of constructed stimulus sequences on held-out video backgrounds.

Retinal parameters are selected and fixed before these assays. Readout directions and thresholds are fitted on 15 complete source videos and frozen for evaluation on five other sources. Supplementary linear logistic readouts select regularization by source-grouped cross-validation within the development set, with standardization computed from the corresponding training folds. Parameter experiments use VisDrone scenes to compare natural histories ending at a target frame with sequences repeatedly presenting that frame, then separately vary response retention and adaptation timescales.

The physiological analysis reuses public ex vivo marmoset retinal ganglion-cell (RGC) recordings. Independent full-field step responses measure response persistence and define sustained-like and transient-like groups. Post-transition spike counts are then used to identify the current stimulus with the preceding stimulus held fixed, or the preceding stimulus with the current stimulus held fixed. Both tasks have four classes. This analysis tests whether response persistence is associated with current-versus-history decoding performance.

Critical strengths
  1. Sections 4.1 and 4.3 use paired inputs to remove direct classification cues. The fork–rejoin assay presents exactly the same continuing video after different histories. Comparisons such as ABC/BAC also fix the last event, preventing classification based only on its identity. Development and evaluation are split by complete source video, and readouts remain fixed after fitting, allowing evaluation on video backgrounds absent from the fitting data. The supplementary experiment combines temporal gains fitted on four stimulus families with known spatial descriptors of two other families to predict their order-readout directions. This tests whether the fitted temporal gains can be used for new patterns while making clear that their spatial information is still supplied.

  2. Section 4.2 separately makes retention or adaptation timescales equal across the two branches while fixing spatial connections and other conditions. The authors compare S/T response differences caused by changes in the current stimulus with those caused by changes in history. Equal retention reduces the angle between these response differences. Equal adaptation preserves a larger angle but reduces subsequent changes in the relative history-related response across the S/T branches. Figure 3 and Figure S2 repeat these checks across parameters, scenes, and history spans, supporting different effects of the two temporal parameters within this model. Section 4.4 additionally finds an association between independently measured response persistence and current-versus-history decoding bias in real RGC populations; groups are not defined directly from the decoding scores being compared.

Critical weaknesses
  1. Some of the main temporal results can be anticipated directly from the equations. In Section 3.2, the fast-response and adaptation states in Equation (4) retain differences caused by earlier inputs; these differences take time to decay after the subsequent inputs become identical. Current activity therefore contains recent history, as expected for this class of dynamics. Earlier and later events receive different decay weights, so generic temporal filters can also distinguish some event orders. Fixing the final event does not remove this capability. Equation (5) directly specifies the retained response under constant input, and the adaptation time constant controls how responses evolve. Changing these parameters is consequently expected to change the relative sizes of current and historical signals. The response-angle and branch-specific history measurements in Figure 3 and Figures S2/S7 remain useful. However, the paper needs to explain whether they further establish when current and past information can be read out together, and which findings go beyond these general filter properties.

    Generic filter controls already support the key order-decoding task. Figure S3C reports terminal-matched order macro AUROC of 0.873 for Raw Two-IIR and 0.683 for RetinaPlex. The joint metric in Table S4 takes the lower AUROC of the current and history tasks. With linear logistic readout, calculating this metric within each video and then averaging gives 1.000 for both Retinal Two-IIR and RetinaPlex; pooling examples from all five videos gives 0.905 and 1.000, respectively. RetinaPlex retains an advantage in the pooled comparison. Scores from individual videos are needed to explain why within-video discrimination is tied but pooled discrimination differs. The existing results do not establish how much of this advantage reflects additional decodable order information or how a fixed classifier scores that information across different videos. Through carefully controlled experiments, the authors show that the model retains information about the order of the constructed stimuli and makes this information decodable. Generic temporal filters also support this capability, and the current analysis does not sufficiently establish what important additional insight the retinal model provides.

  2. Using prior physiology to motivate the model does not validate its particular response rules. Human photoreceptor densities, primate connectivity studies, and sustained/transient responses provide reasonable grounds for the design. Preliminary response checks also show that the selected parameters produce the intended sustained and transient behavior. However, the paper does not directly test whether one-cone input to each S unit and equal-weight six-cone pooling for each T unit, combined with its temporal equations, accurately predict real retinal responses to the same stimuli. Changes observed after varying model parameters directly establish their effects within that model.

    The marmoset analysis in Section 4.4 and Figure S6 supplies another piece of evidence. Sustained-like populations have a larger difference between current-stimulus and preceding-stimulus decoding scores, and the model shows the same direction of difference. This association is useful, but it does not show that the groups exclusively process current or past information, or that transient-like populations achieve higher absolute history-decoding performance. The recordings are not used to validate the model's retention and adaptation parameters, and the analysis does not establish whether alternative temporal mechanisms could produce the same trend. The model mainly analyzes bipolar populations, while the recordings concern RGC output; the existing analysis cannot locate the observed trend in a particular cell layer or identify its governing parameters. Describing the result as an association is appropriate, but agreement in this trend does not establish that real cells follow the model's rules. The added physiological finding is therefore primarily an association between response persistence and relative decoding performance. The paper needs to explain what this relationship adds to the stimulus-history dependence already known from the original study. Together with the expected history retention and the order-decoding ability of generic filters, the current parameter results and association remain insufficient for acceptance.

Questions for authors
  1. Please identify the testable conclusion about the two temporal parameters in Figure 3 and Figures S2/S7 that adds to general knowledge of decaying filters. Specifically, why does equalizing retention reduce the angle between response differences caused by current-stimulus and history changes, while equalizing adaptation preserves a larger angle but reduces subsequent changes in the relative history response across branches? Please provide a brief analysis based on Equations (4)–(5), or use the existing parameter sweeps to explain these effects. Specify the stimuli, parameter ranges, and sampling times for which the conclusions hold. Distinguish results expected from the equations' general gain and decay properties from relationships further established by the experiments.

  2. Please explain the differences across classification methods and score aggregation in Table 1, Figure S3C, and Table S4. Why do Retinal Two-IIR and RetinaPlex tie on average within-video joint AUROC with linear logistic readout in Table S4 but differ after pooling videos? Provide individual-video results or score analyses to determine whether score changes across backgrounds affect the pooled comparison and to explain the remaining advantage. Clarify whether the results establish that RetinaPlex retains order information absent from generic filtering, and which differences depend on the chosen classifier. Also explain which results support conclusions beyond the presence of history information, given that generic filtering decodes earlier order and reports higher performance in Figure S3C. State separately the conclusions supported across readout methods and those that hold only for a particular readout or metric calculation.

  3. Please specify what Section 4.4 and Figure S6 add to the original marmoset study. The original work already showed that current and preceding stimuli jointly affect RGC activity; this paper further groups cells using independent step responses and compares current-versus-preceding decoding. State the new relationship established by this comparison, the results supporting it, and whether it tests particular model parameters. Explain whether the existing recordings permit estimation of retention or adaptation timescales and testing of their relationship with decoding bias. If the present evidence supports only an association between response persistence and decoding bias, state that conclusion and identify the connections, temporal parameters, and cell-layer relationships that remain unvalidated. Clarify which model conclusions the existing physiological results support and which explanations of real retinal mechanisms still require testing.

2 comments
AuthorsOpenReview ↗

Authors' Response: Insights Beyond Generic Filters and Detailed Clarifications

Response to Reviewer TrTb:

We appreciate your technical precision. Your critiques prompted us to add analyses and a new downstream task.

T1. History is expected from generic filters. What is the additional insight?

Response. We agree that retaining a decaying trace is not unique to RetinaPlex. Our claim is about organization and accessible readout within a specified retinal population. The fork–rejoin assays establish joint accessibility, and targeted attenuation reduces readout (recent history: 0.800 to 0.575; present state: 0.875 to 0.700).

However, we provide three important additional insights beyond generic filters:

  1. Functional decoupling of temporal parameters: Generic filters treat decay as lumped. We dissociate retained-response differences (rr, organizing spatial geometry) from adaptation timescales (τa\tau_a, governing temporal reweighting).
  2. Early-hierarchy efficiency and downstream utility: In new Sec. 4.6, a frozen RetinaPlex front end compresses 19 frames into two maps, achieving Macro AUROC 0.729 on a real-video task. This rivals high-dimensional models (e.g., OpenRetina's 64 channels), proving complex temporal reasoning can be supported by low-bandwidth structures established early in the hierarchy.
  3. Biological anchoring of coding bias: Model coding bias finds a concordant association with independently measured RGC temporal phenotypes.

We describe RetinaPlex as an explicitly parameterized and perturbable biological realization, not a unique memory system.

T2. Why does equalizing retention reduce the state–history angle, whereas equalizing adaptation preserves separation but attenuates temporal redistribution?

Response. Eq. (4) has the zero-state transfer function: H(s)=τas+r(1+τfs)(1+τas)H(s) = \frac{\tau_a s + r}{(1 + \tau_f s)(1 + \tau_a s)} The fraction rr sets the steady (DC) gain; τa\tau_a determines a decay pole. Intuitively, rr governs steady-state spatial allocation, while τa\tau_a governs dynamic decay rate. When histories receive identical subsequent drive, the drive cancels and the initial-state difference decays. Whether contrast vectors are non-collinear depends on spatial drives and sampling time.

The experiments quantify the resulting geometry, not a universal theorem. In the VisDrone assay (66.7 ms), the median angle is 56.8° for distinct S/T, 13.7° with equal retention, and 65.7° with equal adaptation. The relative T share of history-residual RMS, Ft=RMSt/(RMSs+RMSt)F_t = \text{RMS}_t/(\text{RMS}_s + \text{RMS}_t), changes from 0.68 to 0.41 (distinct) versus 0.76 to 0.71 (equal adaptation). We distinguish analytically expected effects from stimulus-specific consequences (revised §4.2).

T3. Why do Table 1, Figure S3C and Table S4 give different rankings?

Response. They measure different tasks. Figure S3C uses a terminal-matched order task, where Raw Two-IIR (0.8733) exceeds RetinaPlex (0.6833). We do not use Table 1 to overwrite this stronger generic baseline. Table S4 evaluates logistic readouts of the 128D snapshot. Relative rankings depend on the question, representation, classifier, and aggregation. We clarify this in §4.5.

T4. In Table S4, why do two methods tie within videos but differ when pooled?

Response. Within each of the five held-out videos, both models achieve AUROC 1.0. The pooled difference comes entirely from cross-video comparisons. Among 320 pairs per task, Two-IIR has 9 incorrectly ordered current-state pairs and 38 incorrectly ordered history pairs; RetinaPlex has none. Of those 38 errors, 36 involve positive examples from the "loading" video paired with negatives from other videos. This is a cross-source score ordering difference, not evidence that Two-IIR lacks history information. Revised §4.5 notes that strong generic filters match source-wise discrimination.

T5. What does the marmoset analysis add beyond Krüppel et al.?

Response. We ask whether an independently measured phenotype is associated with the relative balance of current vs. preceding decoding in the same 100–150-ms window. All four recordings show a positive balance difference (0.106–0.206). After matching for amplitude, baseline rate, reliability, and ON/OFF preference, all four effect signs remain positive (three reach Holm-corrected significance; two satisfy both balance and significance criteria). We describe this as phenotype-level concordance in a relative coding statistic, not matched mechanisms.

T6. Do the data establish the anatomical connections or real constants?

Response. No. The one-cone/six-cone connectivity and S/T parameters are anatomically motivated modeling constraints, not recovered biology. The recordings are from RGC output, not paired BC-RGC measurements. Our biological claim is restricted to the phenotype–coding association. This is clarified in the Discussion.

I appreciate the author's response, which addressed most of my concerns. I maintain my positive recommendation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Authors' Common Response & Summary of Revisions

We thank all three reviewers for their careful assessments. We agree that temporal filtering and the dependence of retinal responses on stimulus history are well-established. The contribution we intend to establish is more specific: in a spatially constrained, fixed-parameter retinal population, how jointly accessible present and recent information is organized across sustained/transient pathways; how controlled interventions change that organization; and whether the resulting state is usable by an independent visual task network.

To directly address the concerns regarding AI relevance and generic filters, we have uploaded a substantial revision containing a new downstream real-video experiment (Section 4.6, Figure 5, Table 2). On the UA-DETRAC dataset, a frozen RetinaPlex front end compresses 19 observed frames into only two current S/T maps. These two maps serve as a low-bandwidth preprocessing module for a standard ResNet18-GN decoder, achieving a Macro AUROC of 0.729 for contact prediction. This outperforms the raw current-frame baseline (0.648) while being highly competitive with full 19-frame inputs (0.766).

Furthermore, responding to the suggestion to compare against mature models, we evaluated the bipolar-like intermediate layers of Maheswaranathan et al. (0.717) and OpenRetina (0.777). This comparison highlights a key architectural insight: the efficient organization of present and recent information is already well-established at the early, bipolar-cell stage. RetinaPlex achieves this competitive performance using only 2 channels—compared to OpenRetina's 64 channels—demonstrating that S/T temporal dynamics inherently structure this joint accessibility early in the visual hierarchy, without requiring the downstream complexity or high channel count typically associated with models extracting representations closer to the ganglion cell output. The two-map interface contains 2/192/19 as many stimulus-dependent scalar entries as the 19-frame input.

We have also carefully narrowed our claims wherever the original text could be read as claiming a unique memory mechanism or direct cellular parameter identification. The redlined manuscript now explicitly distinguishes:

  1. The full retinal platform vs. the cone-driven S/T core actually analyzed (Introduction, §3.3).
  2. The 128-dimensional population summary used in the original representational assays vs. the two current spatial maps used by the new trained decoder (§3.3).
  3. The finite-time readout limits and the parameter-intervention scope (§§4.1–4.2).
  4. The marmoset RGC analysis as a "concordant association" rather than a direct mechanism validation, now supported by covariate matching (§4.4).

Note on Supplementary Materials: Several additional supplementary analyses (e.g., continuous parameter sweeps, generic filter baselines) have been completed as research outputs. However, the supplementary pages appended to the current redlined PDF have not yet been fully updated with all final figure/table numbers. We ask reviewers to treat the current text as a scientific response draft regarding these specific supplementary points.

Detailed point-by-point responses to your specific questions are provided in the individual reviewer threads below. We hope these new experimental results and revisions demonstrate the value of RetinaPlex as a stateful, low-bandwidth sensory front end.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.