acceptodds
Under review as a conference paper at ICLR 2027

Probing Selectivity and Tolerance with Optimized Dynamic Stimuli in a Foundation Model of the Mouse Visual Cortex

Abstract

Selectivity and tolerance are fundamental response properties of visual systems. Selectivity describes how concentrated responses are across a stimulus set, whereas tolerance describes response preservation under a specified stimulus transformation. Comparing these properties across visual areas with a common dynamic stimulus protocol remains difficult when reference stimuli are restricted to predefined stimulus families. We analyze the MICrONS-adapted instance of a pretrained video-response foundation model as an in silico hypothesis generator. For each of 104,171 model readouts, we synthesize a most exciting dynamic input (MEDI), a short video optimized in a generative latent space. Compared with static stimuli, MEDIs contain temporal structure, enabling temporal transformations such as reversal and speed resampling. The readout-by-MEDI matrix shows positive within-model-area enrichment, while MEDI-set stimulus selectivity is the highest in model primary visual cortex (V1) and the lowest in model rostrolateral area (RL), with model lateromedial area (LM) and model anterolateral area (AL) intermediate; this pattern is not a simple monotonic progression along the established V1LM/RLAL hierarchy. Transforming each readout's own MEDI reveals transformation-specific tolerance: model RL retains relatively more response under strong spatial transformations, whereas model V1 retains relatively more under slow-down and relatively less under strong speed-up. Across transformations, tolerance does not show a single consistent area ordering. Selectivity and tolerance are negatively associated at selected transformation strengths, but the magnitude of this association depends on transformation and parameter rather than defining a single universal axis. The results revealed profound properties of different populations of readouts of the foundation model and provided experimentally testable predictions for the corresponding visual areas.

Primary area:
applications to neuroscience & cognitive science
Keywords:
visual neuroscience, foundation models, digital twins, stimulus optimization, neural coding, selectivity, transformation tolerance, mouse visual cortex, generative models, in silico experiments

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_e5waOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this work the author analyse an in-silico digital twin of mouse visual cortex (developed by MICrONS Consortium; Wang et al. 2025) by investigating the relationship between neural selectivity (i.e. how big is the pool of stimuli for which a given unit shows substantial responses) and tolerance (i.e. how stable is the unit's response as a given transformation is applied to the provided visual stimulus). The authors leveraged a diffusion-based model to general most exciting dynamic input (MEDI) by optimising the latent codes of an auto-encoder-based generator via gradient descent (the entire setup is fully differentiable) and showed how the resulting optimisation successfully elicited stronger responses then standard alternatives (dynamic Gabors, natural movies, MESI). The generated pool of MEDI stimuli was then use to characterise both sensitivity and tolerance of individual model units: MEDIs were most effective within their target area, selectivity showed the ranking V1 > LM/AL > RL, while tolerance showed various trend depending on the chosen transformation. Finally, selectivity and tolerance were generally anti-correlated and selectivity was stronger in layer L4/L5 neurons than L2/3 neurons.

Critical strengths

The paper presents a comprehensive evaluation of the considered in-silico model, and while the MEDI methodology is not new, its application to the study of the interplay of sensitivity and video-transformations tolerance is interesting. The authors provide good evidence for the success of the optimisation problem and in general offer well-controlled statistical evidence for their results. The paper is in general well written, with clean Figures and a detailed Appendix, furthermore the reproducibility efforts and the nice visualisation website are much appreciated.

Critical weaknesses

While the authors well characterised the model response to MEDI stimuli, it remains unclear how much the derived sensitivity and tolerance trends depend on the chosen methodology. The author compared the methodology to other baselines only to assess MEDI optimisation success, however elevated model responses are by themselves somewhat expected. What would have been perhaps more interesting to assess was how much the derived sensitivity and tolerance ranking depended on the derived stimulus pool. In general, while many results are presented, the author offer little in terms of interpretations, especially for the reported results inconsistent with the known literature, i.e. negative association between selectivity and tolerance. For the latter, the authors point precisely to a difference in the stimulus set, but never test it.

Finally, the proposed technique leveraged a fully differentiable setup to synthesise the MEDI stimulus set. It remains unclear how applicable this technique is to an experimental settings with limits on neuron sampling, optimisation time/number of iterations, noise etc... especially considering the requirement of full-video optimisation. Recent works relied on fast overnight tuning of digital twin (Ding et al. 2026) or turned to gradient-free approaches (Ponce et al. 2023; Tausani et al. ICLR 2026). A comparison of the optimisation convergence time between MEDI and MESI for example would strengthen the claims of applicability to lab experiments.

Questions for authors

The following are questions related to the identified weakness which, if addressed, would further strengthen the paper:

  1. Are the core sensitivity and tolerance results dependent on the derived stimulus set?
  2. Related to (1), In Fig. 9e we see that natural responses, while still less eliciting than MEDI, are the most competing alternatives and arguably escape the stated limitation of optimisation exploiting model-specific errors. How do the results (e.g. tolerance to speed-up, slow-down etc...) look like for naturalistic stimuli?
  3. How do the full-video MEDI optimisations relate to single-frame MESI optimisation (or dynamic Gabors) in terms of number of iterations and computational complexity?
  4. Do the authors estimate the proposed pipeline to be successfully deployed for in-vivo experiments? Would a fine-tuning of the foundational baseline be required?
  5. Some of the explored tolerance transformations (contrast, rotation, translation) are inherently static and can thus be studied in MESI or other single-frame baselines. Does the time-dimension show interesting effects of is the same tolerance profile retrieved in the other conditions?
7 comments
AuthorsOpenReview ↗

Response to Reviewer e5wa #1

Thank you for asking how the findings depend on the stimulus set and how the method could be used experimentally. We added natural-movie and static-reference comparisons, expanded the interpretation, and measured generation time and memory.

Critical Weakness 1 — Stimulus-set dependence

The relationship partly depends on the reference stimuli. We repeated the measurements with preferred natural movies and tested response offsets separately, directly addressing both dependencies you identified.

For a random 2,000-readout cohort (500 per area), we used each target's 60-frame preferred natural video as described in the paper. We evaluated all 2,000 references against all targets. A reference's source area is the area of the readout that selected it. We used the same definitions of selectivity and tolerance and the same transformation settings as for MEDIs.

MeasureResult consistent with MEDI referencesResult inconsistent with MEDI references
Within-area preferenceV1, RL, and AL preferred their own source groups.LM preferred another source group.
SelectivityV1 was highest; medians were 0.570, 0.504, 0.467, and 0.445 in V1, LM, RL, and AL./
Translation, speed-up, slow-downV1 was lowest for translation and speed-up and highest for slow-down; all nine V1-versus-other-area comparisons passed Holm correction./
RotationAL had the highest median.the AL-RL difference was unresolved after correction (p=0.403p=0.403).
Selectivity-tolerance associationsContrast, rotation, translation, and slow-down correlations were negative in every area.Speed-up correlations were near zero (ρ=−0.063\rho=-0.063 to 0.0430.043), as was V1 reversal (0.0060.006).

An interesting finding is that, speed-up increased the median response in every area, giving ratios of 1.049-1.147. The originals were selected among untransformed natural clips, so a sped-up version could elicit a stronger response. V1 had the smallest increase.

Different areas may select similar or identical natural clips, making the source groups less distinct and offering one possible explanation for weaker area specificity. Several temporal associations also changed when natural references replaced MEDIs (Appendix F.4, Figure 25).

Therefore, these differences show that within-area preference and the selectivity-tolerance relationship depend on the stimulus set and reference. The MEDI framework complements predefined stimulus families by identifying strongly driving dynamic references for individual readouts, and applying the same selectivity and transformation protocol across areas.

Critical Weakness 2 — Generation cost and deployment

MEDI generation took about one minute per target on the measured GPU. We generated all four stimulus types for ten readouts on an NVIDIA RTX 4090 at the settings in Appendices B.1-B.2. Entries are medians (minimum-maximum):

StimulusTime, sPeak memory, GiB
MEDI59.19 (56.37-65.82)2.91 (2.90-2.97)
MESI3.03 (2.83-3.44)3.72 (3.72-3.80)
Static Gabor24.65 (23.94-25.39)11.02 (10.97-11.20)
Dynamic Gabor24.64 (23.81-25.89)11.02 (10.97-11.20)

Timing excludes loading and saving; peak tensor allocation includes resident models and the decoder when used (Appendix F.8).

We optimize an autoencoder's latent representation by gradient descent. MEDI generation uses eight overlapping 15-frame windows, with 8-frame overlap and 20 updates per window, before stitching and trimming to 60 frames. MESIs use 10 iterations; both Gabors use 30. The table reports the costs of these procedures at their stated settings.

AuthorsOpenReview ↗

Response to Reviewer e5wa #2

Question 1-2 — Stimulus-set dependence and natural-reference transformations

The core sensitivity and tolerance results are partly dependent on the derived stimulus set. See Critical Weakness 1 and Appendix F.4, Figure 25: under natural-reference stimulus set, several profiles persist, but LM preference and some temporal associations change.

Question 3 — Iterations and computation

MEDI generation uses 20 updates per overlapping window; MESI and dynamic-Gabor generation use 10 and 30 iterations, respectively. Their parameters and the work per update differ, so measured wall time and peak memory provide the practical comparison.

Appendix F.8 reports time and memory at these settings. Pilot experiments would establish an appropriate iteration schedule for a new animal.

Question 4 — In-vivo deployment

Deployment requires fine-tuning the animal-specific components of the pretrained model on recorded responses. The pretrained core supplies visual features while the trained components adapt predictions to the recorded animal. The cited overnight digital-twin work provides a precedent for model adaptation, while gradient-free methods offer a complementary approach to stimulus search.

At the measured generation rate, 500-1,000 MEDIs would require approximately 8-16 GPU-hours. With ideal parallel scaling on four comparable GPUs, synthesis alone would take about 2-4 hours; model training, data transfer, and recording add to that time.

Our planned experiment would record natural-movie responses, train the animal-specific model and generate stimuli overnight, then test optimized and conventional stimuli on subsequent days.

Pilot recordings will establish prediction quality, optimization settings, repetition counts, and total time consumption. The Discussion describes this planned validation, and Appendix F.8 provides the measured synthesis costs.

Question 5 — Static spatial tolerance

Yes. We tested paired MESIs and MEDIs in the random 500-per-area cohort. MESIs were repeated images; each pair shared a rotation center. Retention used each family's own untransformed response as denominator.

MESIs also showed the lowest translation retention in V1 and the highest rotation retention in AL. At 48 pixels, medians were 0.039, 0.091, 0.131, and 0.159 in V1, LM, RL, and AL; at 90∘90^\circ, they were 0.097, 0.148, 0.149, and 0.189 (Appendix F.7, Figure 28).

The close spatial similarity between MESIs and MEDI mean frames offers a plausible explanation for this agreement. The additional value of the video is its temporal sequence: in the separate 100-readout ablation, repeating the strongest individual frame retained a median 35.6% of the MEDI response and shuffling retained 49.6%. MEDIs therefore share spatial structure with MESIs while adding response-driving temporal structure and enabling reversal and playback-speed interventions (Appendix F.6, Figure 27). We now make this division of roles explicit in the Discussion.

I thank the authors for the rebuttal and the new analyses responding to my questions, in particular the natural-movie reference analysis (now Appendix F.4), the comparison with MESIs on tolerance, and the report of the computational costs relevant for in-vivo deployment. I believe these additions have improved the paper; however, as I understand them, they have also reinforced my original concerns. In particular:

  1. The new experiments show that the selectivity–tolerance association depends strongly on the chosen stimulus set: with preferred natural movies, the within-area correlations roughly halve for contrast and translation and largely vanish for reversal and speed-up. Moreover, retention above one under speed-up suggests that retention partly reflects how far the reference stimulus is from the readout’s optimum. Since MEDIs are by construction local optima of the model response, a readout with a narrowly peaked response would appear both selective and intolerant. This makes it hard to tell whether the negative association reflects real properties of the model’s visual areas (and might generalise to visual-cortex recordings), or is simply a consequence of measuring tolerance around a response maximum. A dose-response check (e.g., using partially optimised MEDIs) could help separate the two.
  2. The MEDI–MESI comparison shows that MESIs reproduce the spatial tolerance profiles (in Fig. 28 the dashed and solid curves closely track each other). For the tolerance analysis, the distinctive contribution of MEDIs therefore lies in the temporal manipulations. While the area-level temporal profiles are robust, the temporal selectivity–tolerance associations (reversal, speed-up) are precisely the ones that weaken across response models or depend on the stimulus set.

For these reasons I keep my original score: the new analyses improved the work, but they leave these original concerns open.

Reviewer_5xC9OpenReview ↗
Rating
3(+1)
Confidence
1(−3)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper uses a pretrained video-response foundation model adapted to the MICrONS mouse to investigate selectivity and transformation tolerance across visual areas. The authors optimize a most exciting dynamic input (MEDI) for each of 104,171 model readouts. They measure stimulus selectivity using area-balanced lifetime sparseness and tolerance using the response retained after transforming each readout’s own MEDI.

The main findings include within-area response enrichment, area-dependent selectivity without a simple monotonic hierarchy, transformation-dependent tolerance profiles, negative selectivity–tolerance associations at selected transformation strengths, and lower median selectivity in L2/3 than in L4 and L5. A further analysis distinguishes high reversal retention accompanied by small versus large changes in the stimulus.

Critical strengths
  • The the sturcture of this paper is very clear and easy to follow.
  • Using MEDI to study relationship between selectivity and tolerance is interesting.
  • This paper offers an useful framework for studying dynamic response properties. Optimizing neuron-specific videos provides strongly driving reference stimuli and enables temporal interventions, including reversal and playback-speed changes.
Critical weaknesses

Major weaknesses

  • The MEDIs are optimized and evaluated based on the same model, significantly weakening the claim in section 4.1. The authors acknowledge this limitation and frame the study as hypothesis generation. However, the significance of the empirical contribution would be stronger with an additional control evaluation model using an independently fitted predictor.
  • The effect of receptive fields is overlooked, entangling the spatial and feature selectivity/tolerance. The MEDIs are spatially localized, and the foundation model in the paper uses neuron-specific readout positions. As a result, a target readout's response to another readout's MEDI can depend on receptive field as well as feature preference. It is not very fair to directly input the MEDI of a neuron in an early area to the readout of a neuron in the later area when their receptive fields differ significantly, weakening claims in some sections. This problem also applies to translation transformation. The translation uses a fixed set of absolute pixel numbers and don't take into account the receptive field size. The control of receptive field or the evidence showing that the receptive filed doesn't affect cross-readout results is essential.
  • Soundness concern: some descriptions in Section 4.3 appear inconsistent with Figure 4. The text describes an increasing speed-up profile across the displayed areas and states LM < RL < AL for temporal reversal, whereas panels e and d appear to show a decrease from RL to AL.

Minor weaknesses

  • In Figure 2b and Figure 9, it is hard to distinguish areas.
  • The aggregation procedure needs a consistent description. Figure 4’s caption refers to medians across repeated entries, whereas Appendices B.3–B.4 describe averaging repeated entries. The distinction between aggregation across repeats, transformation parameters, and readouts should be explicit.
Questions for authors
  • Were MEDIs constrained or normalized to a common contrast during optimization? If their baseline contrasts differ, applying the same multiplier may probe different portions of each readout’s contrast–response curve, potentially affecting response retention.
  • Additional papers that should be cited:
  1. Luo, A., Henderson, M., Wehbe, L., & Tarr, M. (2023). Brain diffusion for visual exploration: Cortical discovery using large scale generative models. Advances in Neural Information Processing Systems, 36, 75740-75781.
  2. García Cerdas, D., Sartzetaki, C., Petersen, M., Roig, G., Mettes, P., & Groen, I. (2025, May). BrainACTIV: Identifying visuo-semantic properties driving cortical selectivity using diffusion-based image manipulation. In International Conference on Learning Representations (Vol. 2025, pp. 9040-9075).

Overall assessment

I currently lean toward rejection. The framework is useful, the empirical findings are potentially informative, and several controls are thoughtfully designed. The receptive field effect is my major concern. When browsing the website provided in the paper, most receptive fields don't appear to be very dfferent. But some explicit analyses and evidence is required. Focused clarification and controls addressing these issues could substantially improve my assessment.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 5xC9 #1

Thank you for explaining the receptive-field concern and the analyses needed to address it. We will address each points in order.

Major Weakness 1 — Shared optimization and evaluation model

The four main claims (see General response to all reviewers) persist in three separately trained models. We selected the 500 readouts with the highest ccmax⁡cc_{\max} in each area, giving 2,000 targets. Here, ccmax⁡cc_{\max} estimates recording reliability from repeated presentations of the same movie. Two models use the original readout architecture with different training seeds, and the third uses a two-block causal Transformer readout, all retaining the pretrained core. Checkpoints were selected by maximum median ccnormcc_{\mathrm{norm}}, the prediction accuracy.

The original model generated every stimulus, then each new model evaluated these unchanged stimuli. We compared the MEDI response with the response to each baseline stimulus predicted by each new model, taking the median ratio across 2,000 comparisons for each condition. As shown in figure 21, all five MEDI-baseline comparisons were significant in each model, both pooled and within each area (two-sided Wilcoxon signed-rank tests; all Holm-adjusted p<0.01p<0.01):

StimulusSeed 2042Seed 3042Transformer readout
MESI2.722.692.94
Static Gabor3.443.243.57
Dynamic Gabor1.991.891.69
Drifting grating2.031.911.85
Preferred natural movie1.461.421.18

As shown in figure 22, all three models retained within-area preference and highest selectivity in V1. V1 had the lowest median retention under translation and speed-up and the highest under slow-down; AL had the highest rotation median. All 12 comparisons underlying the four tolerance directions passed Holm correction except seed 3042 V1-LM speed-up (adjusted p=0.053p=0.053). Of the 24 selectivity-tolerance associations, almost all were negative in all three models. Specifically, V1 speed-up was near zero (ρ=−0.008\rho=-0.008, 0.0600.060, 0.0460.046), as was LM speed-up in the Transformer (−0.052-0.052; Methods 3.4; Appendix F.2).

Therefore, we showed that the four main claims persist in three separately trained models.

Major Weakness 2 — Spatial confounds

Spatial position. We used released readout positions as receptive-field center proxies. For each target, we matched source-target distances across areas by selecting equal numbers of source MEDIs in each distance bin. We compared these pools with full pools and equally small random pools to separate distance matching from the effect of using fewer stimuli. Matched pools contained 77-150 stimuli per source area.

Across two 4,000-target samples (1,000 per area) and five matching repetitions, within-area preference and highest V1 selectivity remained. All 48 selectivity-tolerance associations across two sampling seeds were negative (ρ=−0.828\rho=-0.828 to −0.143-0.143; Appendix F.11, Figures 31-33).

Translation scale. For a random 2,000-readout cohort (500 per area), we measured each readout's receptive field approximately from its MEDI's temporal range map: the intensity range at each pixel. These values, together with horizontal and vertical spatial standard deviations, defined an axis-aligned ellipse. We then scaled horizontal and vertical shifts by the corresponding semiaxis. As predicted, V1 had the lowest median retention at all four tested fractions, 0.25, 0.5, 1, and 2. At one semiaxis, medians were 0.119, 0.156, 0.158, and 0.175 in V1, LM, RL, and AL.

We also matched both MEDI spatial scales across areas, retaining 296 readouts per area. V1 remained the lowest median retention in all 100 selections, both at the original 48-pixel shift and at one semiaxis. The V1 difference therefore persists after controlling the receptive field effects (Appendix F.12, Figures 34-35).

Rotation geometry. For a random 2,000-readout cohort (500 per area), we measured elongation and the image change caused by rotation. Image change was measured by the squared pixel difference divided by the original video's squared deviations from frame means. Elongation had little association with retention (∣ρ∣≤0.043|\rho|\leq0.043); larger image changes were associated with lower retention (ρ=−0.396\rho=-0.396 to −0.315-0.315). Matching both measures still left AL numerically highest, with an unresolved AL-RL difference: 0.0118, 95% interval [−0.0024,0.0252][-0.0024,0.0252], adjusted p=0.262p=0.262. Thus, elongation confounds minimally rotation tolerance, whereas internal pattern changes help interpret moderately (Appendix F.13, Figure 36).

AuthorsOpenReview ↗

Response to Reviewer 5xC9 #2

Major Weakness 3 — Incorrect area orderings

We removed the unsupported LM-RL-AL orderings for speed-up and reversal. Section 4.3 now reports only the four supported median directions: V1 lowest for translation and speed-up, highest for slow-down; AL highest for rotation at selected settings.

Minor Weakness 1 — Area colors

The revised panels separate areas by position and color. They show individual ratios, medians, and interquartile ranges; one denotes equal responses. Values above 20 appear as triangles; all statistics use uncapped ratios.

Minor Weakness 2 — Aggregation order

At each parameter value, we first average repeated entries within each readout, including the directions for rotation and translation. We then take the median across readouts within each area. Figure 4 shows one selected setting per transformation; Figure 10 shows the separate parameter values (Methods 3.3; Appendices B.3-B.4).

Question 1 — Starting contrast

Contrast was free during optimization under the common objective and regularization settings. We added a control that equalizes starting contrast before applying the contrast multipliers.

For a random 2,000-readout cohort (500 per area), we rescaled MEDIs to global root-mean-square (RMS) contrast 0.05 while preserving each mean, then applied framewise multipliers. Each denominator was the RMS-matched original response.

At 0.25×0.25\times contrast, all four selectivity-tolerance associations remained negative (ρ=−0.680\rho=-0.680, −0.799-0.799, −0.813-0.813, −0.745-0.745), using selectivity from the original pool. At increased contrast, the associations became positive. Therefore, the negative association at the selected setting persists after matching starting contrast (Appendix F.5, Figure 26).

Question 2 — Related work

We added both papers to Related Work and the bibliography:

  • Luo et al. (2023), BrainDiVE: synthesizes images to explore visual-region preferences using predicted fMRI responses.
  • García Cerdas et al. (2025), BrainACTIV: modifies reference images to increase or decrease predicted regional responses while preserving image structure.

We now place our dynamic-reference analysis alongside these generative approaches.

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

The paper studies selectivity and tolerance in four mouse visual areas: V1, LM, RL, and AL. To do so, the authors use a publicly available MICrONS foundation model and synthesize a most exciting dynamic input (MEDI) for each readout using regularized optimization in a pretrained autoencoder’s latent space. Selectivity is measured using area-balanced lifetime sparseness across the MEDI library, while tolerance is measured as response retention after transforming each readout’s own MEDI.

Selectivity is highest in model V1 and lowest in model RL, rather than following a monotonic progression along the proposed visual hierarchy. Tolerance profiles depend on the transformation, with no single area ordering across conditions. Within each area, more selective readouts generally show lower tolerance at the selected transformation strengths, although the association varies across transformations. These findings provide testable predictions.

Critical strengths
  1. The dynamic stimulus framework are a neat way to move into temporal transformation which goes beyond usual spatial tolerance studies.
  2. The protocol provides a good way to have systematic comparisons.
  3. I find very useful the controls that were designed and promptly acknowledging limitations.
Critical weaknesses
  1. Sensitivity to stimulus regularization may be insufficiently tested. MEDIs are generated under spatial an temporal smoothness constraints which may favor particular stimulus properties and affect comparisons across visual areas. A control, checking whether the selectivity-tolerance association survives regularization strength or controlled background perturbations
  2. The link between selectivity and tolerance may partly reflect how they are measured. Tolerance is measured by changing a stimulus optimized to strongly activate each readout. Readouts that respond strongly to only a narrow range of stimuli may also lose more response when their optimized stimulus is changed. Differences in baseline activity could further contribute to this association. Excluding each readout’s own MEDI from the selectivity calculation does not address these possibilities. Testing whether the relationship persists with natural-movie reference stimuli and after accounting for baseline activity would strengthen the interpretation.
Questions for authors
  1. Could the smooth backgrounds and spatial regularization used to generate MEDIs contribute to the higher selectivity observed in V1? It would be useful to test whether adding controlled background noise, or varying the regularization strength, changes the selectivity–tolerance relationship and the ordering across areas.”
3 comments
AuthorsOpenReview ↗

Response to Reviewer nHaQ #1

Thank you for identifying regularization, stimulus references, and response baselines as potential sources of the relationship. We tested each directly.

Critical Weakness 1 — Regularization and background noise

V1 remained most selective and all selectivity-tolerance associations remained negative under the regularization and background controls. For a random 2,000-readout cohort (500 per area), we experimented eight more settings apart from the original optimization protocol: a second optimization seed, half or double the spatial penalty, half or double both temporal penalties, noisier initialization (SD 0.10 versus 0.02), and shorter or longer windows (10 frames/5-frame overlap and 20/10, versus the original 15/8).

We also tested two background-perturbation conditions. For each MEDI, we identified the smallest spatial footprint, containing 98% of its temporally averaged squared deviations from the frame means, and expanded this protected region by five pixels. We added Gaussian noise (SD 0.01 or 0.03) only outside this region, leaving the protected pixels unchanged (Appendix F.1).

Across all 11 conditions (including the original protocol), each area still preferred MEDIs from its own source group and V1 remained most selective. All tolerance-related claims were also retained: V1 had the lowest translation and speed-up tolerance and the highest slow-down tolerance; AL had the highest rotation tolerance. Every selectivity-tolerance association estimate remained negative (Appendix F.1, Figures 18-20).

AuthorsOpenReview ↗

Response to Reviewer nHaQ #2

Critical Weakness 2 — Reference and response baseline

The relationship partly depends on the reference stimuli. We repeated the measurements with preferred natural movies and tested response offsets separately, directly addressing both dependencies you identified.

For the same random cohort, we used each target's 60-frame preferred natural video as described in the paper. We evaluated all 2,000 references against all targets. A reference's source area is the area of the readout that selected it. We used the same definitions of selectivity and tolerance and the same transformation settings as for MEDIs.

MeasureResult consistent with MEDI referencesResult inconsistent with MEDI references
Within-area preferenceV1, RL, and AL preferred their own source groups.LM preferred another source group.
SelectivityV1 was highest; medians were 0.570, 0.504, 0.467, and 0.445 in V1, LM, RL, and AL./
Translation, speed-up, slow-downV1 was lowest for translation and speed-up and highest for slow-down; all nine V1-versus-other-area comparisons passed Holm correction./
RotationAL had the highest median.the AL-RL difference was unresolved after correction (p=0.403p=0.403).
Selectivity-tolerance associationsContrast, rotation, translation, and slow-down correlations were negative in every area.Speed-up correlations were near zero (ρ=−0.063\rho=-0.063 to 0.0430.043), as was V1 reversal (0.0060.006).

An interesting finding is that, speed-up increased the median response in every area, giving ratios of 1.049-1.147. The originals were selected among untransformed natural clips, so a sped-up version could elicit a stronger response. V1 had the smallest increase.

Different areas may select similar or identical natural clips, making the source groups less distinct and offering one possible explanation for weaker area specificity. Several temporal associations also changed when natural references replaced MEDIs (Appendix F.4, Figure 25).

Response offsets. We agree that baseline activity could contribute to selectivity and tolerance. Across the MEDI-readout responses, 44.6% were below the gray-screen response and 48.0% below the black-screen response. Subtracting either screen response and clipping negative values at zero would collapse all below-screen responses to zero, discarding differences among them. Rolls and Tovee (1995) noted this limitation of baseline-subtracted, rectified responses when stimuli suppress firing below the spontaneous rate, and suggested using the lowest stimulus-evoked response as an alternative reference. Following this approach, we recalculated both selectivity and tolerance relative to each readout's minimum response across the shared MEDI set.

For the same random sample of 2,000 readouts (500 per area), we defined mim_i as each readout's minimum response across the 2,000 MEDIs. We subtracted mim_i before computing selectivity and from both responses in the tolerance ratio, [ri(transformed)−mi]/[ri(original)−mi][r_i(\mathrm{transformed})-m_i]/[r_i(\mathrm{original})-m_i]. Within-area preference, highest V1 selectivity, and all four principal tolerance directions persisted. All 24 within-area selectivity-tolerance associations remained negative (ρ=−0.850\rho=-0.850 to −0.189-0.189; all Benjamini-Hochberg-adjusted p<0.001p<0.001). Thus, these findings persist when both metrics use this alternative response reference (Appendix F.10, Figure 30).

Question — Could regularization explain V1 selectivity?

V1 remained most selective under half/double spatial penalties and both background-noise levels; see Critical Weakness 1 and Appendix F.1, Figures 17-20.

AuthorsOpenReview ↗

Clarification on stimulus-set dependence and the scope of our conclusions

We would like to clarify one point in our previous response, as we realize that the statement that the selectivity-tolerance relationship partly depends on the reference stimuli could be interpreted more broadly than we intended.

Our point is not that the MEDI-based relationship is unreliable because it is weaker under natural-movie references. Rather, selectivity and tolerance are measured relative to a particular reference ensemble, and changing that ensemble changes the question being examined. The MEDI analysis asks how readouts differ in their responses across a shared set of strongly driving, readout-specific dynamic references, and how their responses change when these references are transformed. The natural-movie analysis provides a complementary measurement using a predefined, naturalistic reference ensemble.

We consider the differences between these two analyses informative. They show that the selectivity-tolerance relationship is not a universal property independent of the stimuli used to measure it. At the same time, this does not invalidate the relationship observed with MEDIs. The two analyses characterize responses under different reference conditions, and we do not expect every association to be identical across them. Our conclusions are therefore specifically about the relationships measured with the MEDI ensemble, rather than a claim that the same relationships must hold for every possible stimulus ensemble.

We have also performed an additional experiment that addresses a separate concern: whether the negative association arises simply because MEDIs are optimized to maximize responses. Using the same 2,000 readouts (500 per area), we regenerated MEDIs with 10 rather than 20 optimization updates per window, keeping the other settings and learning-rate schedule unchanged. The resulting stimuli elicited substantially lower responses, with median paired half/full response ratios of 0.44-0.49 across areas. Nevertheless, all 24 within-area selectivity-tolerance correlations remained negative. Within-area preference, highest V1 selectivity, and the four principal area-level tolerance directions also persisted (Appendix F.1, Figures 18-20).

We see these two comparisons as addressing different questions: the natural-movie analysis examines dependence on the reference ensemble, whereas the half-iteration experiment examines dependence on the degree of MEDI optimization. Together, they support a more specific interpretation: the observed relationship depends on the reference ensemble, but is not restricted to fully optimized MEDIs.

We wanted to make this distinction explicit because we believe stimulus-set dependence is an important part of the scientific finding, not a reason to dismiss the MEDI-based results. We have kept our conclusions within the scope supported by these experiments and do not claim a universal selectivity-tolerance relationship across all stimulus ensembles.

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

This paper uses a foundation model of mouse visual cortical neural responses, trained on the very large MICrONS corpus, as an in-silico hypothesis generator. In particular, their model-dissection focuses on two properties of visual neurons: stimulus selectivity and tolerance. Authors note that previous published research has focused on simplistic synthetic stimuli to study these effects, although it is now well-known that most exciting inputs (MEIs) derived from digital twins, elicit stronger responses. In addition, authors extend static MEIs to the dynamic case, by optimizing a video stimulus rather than a single image, showing higher drives in their model neurons, and allowing temporal manipulations to assess selectivity and tolerance. Both are finally compared against each other and across areas/layers.

Critical strengths

I really appreciate the effort of optimizing dynamic MEIs, which is indeed not covered enough in the literature -the authors made a great job covering the few approaches that do. Leveraging the temporal dimension to elicit even stronger responses has the potential of telling us more about how visual neurons compute, not only along the spatial axis.

The motivation of the paper is well-founded in my opinion, and I would say that the core idea of the paper -ie, exploring selectivity/tolerance along the temporal dimension and around “optimal” inputs- constitutes the paper’s main strength.

At a macroscopic level, the paper progresses quite naturally even with some downsides (see weaknesses below); I also appreciate how careful and honest the authors are with their results, by not overselling them, but that also creates discrepancies between the abstract and the conclusion.

The methods are large-scale, the information is there -although the form could be improved- and the whole study seems reproducible. The companion website is all the more appreciated, that videos are staple for this paper.

As a whole this submission constitutes a substantial engineering effort, and tries to answer a legitimate neuroscience question in silico…

Critical weaknesses

…However, some weaknesses I find too important prevent me from recommending this article for publication in the present state. On the content:

  • It is perhaps a good sanity check that optimized MEDIs outperform other baseline stimuli, but I do not find it surprising, since the training and evaluation models are the same. This points towards a bigger weakness of the paper that prevents me from holding these results as really significant: I did not see any assessment of the robustness of the results with the model seed -they could be idiosyncratic to this particular model instance. Checking them on one or two other seeds would be reassuring. In general, to ensure MEIs are actually meaningful, one optimizes them on a model ensemble; one can even split the ensemble into validation and test, such that the final MEIs have nether seen the test models. The authors present ridiculously low p-values (eg, 1e-100), but that is not easy to obtain with such a high population (tens of thousands) and with a single model instance. Similarly, I wonder whether the results hold from one animal to another.

  • Another potential confound with the reversed stimuli. The "response" is the temporal firing rate over the whole stimulus presentation, discarding the temporal dimension entirely. So original and reversed stims could elicit very temporally different responses, which could still have the same temporal average, and therefore a high retention... Worse: a fully linear or memoryless system would just output a temporally reversed response, and there, retention would be maximal, though responses would be qualitatively different. In my personal opinion, taking a video model to leverage the temporal dimension in the inputs (stimuli), but discarding it in the output (responses) by averaging is a bit of a pity, though I recognize it is defensible. But the authors should defend it more.

  • Related to the above: to me, optimizing video MEIs properly could be a contribution in itself; but there are many pitfalls and therefore the authors must make show convincingly that their method brings benefits compared to static ones in the first place. Which means, using similar levels of validation as for static MEIs. I’m not asking to validate them in vivo, but more checks: are MEDIs related to MESIs (from the visualization website, it looks like), to what extent, etc?

  • My other big concern is that the authors do not seem to try and interpret why their results, are the way they are. For example, from figure 4c, it seems that higher-order visual areas show more retention with translation: I bet that this could simply be explained by larger receptive fields. Similarly, retention to rotation will depend on how elongated (ie, not a round, blob) each MEDI is. A last example: lifetime sparseness is invariant to scaling but very sensitive to adding a bias. So selectivity, and therefore the tradeoff that follows with retention, may just be the product of high baseline firing rate.

  • Choosing 200 neurons (per area) randomly is not standard and poses questions. In general, one prefers a better choice heuristic than chance; for example, best predicted units.

Note that these do not constitute an exhaustive list. In terms computational neuroscience, the methods -if perfected- constitute a very welcome contribution. In terms of ML, I do not see much novelty.

On the form:

  • This might be personal, but I find the overall writing to be at times very convoluted and hard to follow. For example, “MEDI-set”, “enrichment matrix” (?) are heavy, and the sentence “[…] whether the area-level selectivity pattern was accompanied by systematic differences across cortical layers within each model area.” could be simplified -these are just examples.

  • The writing hedges way too much, in ways that are the hallmark of certain LLMs such as Claude, from my personal experience. A few examples: “These parallels provide qualitative correspondence rather than validation of the model-derived area patterns.”; “the model may depend more on framewise texture or shape than on temporal order.”; “it is an enrichment summary, not a connectivity map.”; and so on… If anything, that dilutes the message and makes everything less clear, in my personal opinion.

  • More importantly, the whole paper reads a bit like an internal log -a succession of paragraphs, cf., conclusion and discussion-, without enough effort to link the different parts together to make the ensemble more accessible to readers.

  • Many important “details” such as how animal behavior is handled/given to the foundation model are completely left out needs to be fished in the appendix, breaking the flow of the paper. How are frames interpolated in the speedup and slowdown conditions?

  • The speed results (V1 retains more under slow-down and less under speed-up, while RL and AL do the opposite) line up with classic findings that AL and RL prefer higher temporal frequencies and speeds than V1. Making these links explicit would be both a validation and a more honest framing: several 'findings' are rediscoveries of known tuning differences seen through a new metric. That’s fine, but I consider it should be said.

  • Some figures such as figure 2b and 9 have bad coloring -too bland-, making them hard to read. Some figures could be more informative and serve the content better, such as tolerance vs selectivity.

  • Fig. 3b’s caption defines “n.s.” without any being found in the plot

Again, I cannot go on everything here for the sake of brevity, but all in all, this paper feels heavily written by AI, which by itself is not a problem if proper rephrasing and manual writing is given to ensure that all readers stay onboard. Thus, I would I recommend a substantial re-writing and rephrasing.

Questions for authors
  • Adress possible confounds that could come mechanically from the readout parameters (see weaknesses above)
  • Examine a tolerance vs selectivity scatter, its shape, etc. rather than just one spearman coefficient (figure 5). There might be much more to say.
  • Check robustness across model seeds? Animals? Do the results hold up?
  • Clarify your random choice of 200 neurons, and why not another heuristic.
  • Adapt the writing
  • Are MEDIs temporally stationary? More quantitative analysis on those would be welcome.
11 comments
AuthorsOpenReview ↗

Response to Reviewer UbuW #1

Thank you for recognizing the value of dynamic optimization and for identifying the validation and interpretation it needs. We address those concerns and the presentation points in your original order.

Critical Weakness 1 — Model and animal robustness

The four main claims (see General response to all reviewers) persist in three separately trained models. We selected the 500 readouts with the highest ccmax⁡cc_{\max} in each area, giving 2,000 targets. Here, ccmax⁡cc_{\max} estimates recording reliability from repeated presentations of the same movie. Two models use the original readout architecture with different training seeds, and the third uses a two-block causal Transformer readout, all retaining the pretrained core. Checkpoints were selected by maximum median ccnormcc_{\mathrm{norm}}, the prediction accuracy.

The original model generated every stimulus, then each new model evaluated these unchanged stimuli. We compared the MEDI response with the response to each baseline stimulus predicted by each new model, taking the median ratio across 2,000 comparisons for each condition. As shown in figure 21, all five MEDI-baseline comparisons were significant in each model, both pooled and within each area (two-sided Wilcoxon signed-rank tests; all Holm-adjusted p<0.01p<0.01):

StimulusSeed 2042Seed 3042Transformer readout
MESI2.722.692.94
Static Gabor3.443.243.57
Dynamic Gabor1.991.891.69
Drifting grating2.031.911.85
Preferred natural movie1.461.421.18

As shown in figure 22, all three models retained within-area preference and highest selectivity in V1. V1 had the lowest median retention under translation and speed-up and the highest under slow-down; AL had the highest rotation median. All 12 comparisons underlying the four tolerance directions passed Holm correction except seed 3042 V1-LM speed-up (adjusted p=0.053p=0.053). Of the 24 selectivity-tolerance associations, almost all were negative in all three models. Specifically, V1 speed-up was near zero (ρ=−0.008\rho=-0.008, 0.0600.060, 0.0460.046), as was LM speed-up in the Transformer (−0.052-0.052; Methods 3.4; Appendix F.2).

Therefore, we showed that the four main claims persist in three separately trained models.

The response advantage and negative population-level associations also persist in a second mouse. For mouse 27203 from Wang et al., we trained two models using the original architecture, retaining the pretrained core. We selected the 2,000 highest-ccmax⁡cc_{\max} readouts. The seed-0 model generated the stimuli; the seed-1042 model evaluated them unchanged. Median MEDI/baseline ratios were 3.36, 6.29, 6.10, 6.28, and 2.72 for MESIs, static Gabors, dynamic Gabors, gratings, and natural movies. All five MEDI-baseline paired comparisons were significant (two-sided Wilcoxon signed-rank tests; all Holm-adjusted p<0.001p<0.001). All six population-level correlations were negative (ρ=−0.800\rho=-0.800 to −0.249-0.249), with readout-bootstrap intervals below zero. Area labels were unavailable, so this test concerns the pooled population (Appendix F.3, Figures 23-24).

To address concerns regarding pp values, we now report effect sizes, response ratios, median differences, and readout-bootstrap intervals. These intervals describe variation among model readouts (Appendices B.3-B.4).

Critical Weakness 2 — Mean response versus response trajectory

You are right that equal mean responses can hide different response trajectories. We retain the mean to measure response magnitude and now compare the trajectories separately.

For a random 2,000-readout cohort (500 per area), we presented each 60-frame MEDI and its exact reversal. We correlated each readout's original MEDI response trace with its reversed-MEDI response trace, and with its reversed-MEDI reversed response trace. The table reports area medians of Pearson correlations:

Correlation with original MEDI response traceV1LMRLAL
Reversed-MEDI response trace0.3790.3730.3090.371
Reversed-MEDI reversed response trace−0.278−0.322−0.297−0.389

A memoryless model would show correlation of 1 in the second row, which is not the case. The negative median correlations show that the same images elicited different responses after different visual histories. Mean retention and trajectory similarity therefore answer different questions.

The Discussion now distinguishes response magnitude from temporal pattern (Appendix F.9, Figure 29).

AuthorsOpenReview ↗

Response to Reviewer UbuW #2

Critical Weakness 3 — MEDIs versus static references

MEDIs often share spatial structure with MESIs, but temporal sequence adds response. In 100 randomly selected readouts, 25 per area, the median pixel correlation between each MESI and its paired MEDI mean frame was 0.911.

For each readout we divided each control response by its original MEDI response, then took medians across readouts:

ControlMedian response ratioReadouts with ratio < 1
Repeated mean-MEDI frame0.199100%
Repeated MEDI most exciting frame0.35698%
Shuffled MEDI frames0.496100%
Repeated MESI frame0.29798%

Each static image was repeated for 60 frames. The shuffled control uses the median response across five permutations per readout; it preserves the frames while disrupting order and continuity.

Both static controls and shuffling reduced median response, supporting contributions from temporal sequence (Appendix F.6, Figure 27).

Critical Weakness 4 — Spatial and response-offset confounds

Spatial position. We used released readout positions as receptive-field center proxies. For each target, we matched source-target distances across areas by selecting equal numbers of source MEDIs in each distance bin. We compared these pools with full pools and equally small random pools to separate distance matching from the effect of using fewer stimuli. Matched pools contained 77-150 stimuli per source area.

Across two 4,000-target samples (1,000 per area) and five matching repetitions, within-area preference and highest V1 selectivity remained. All 48 selectivity-tolerance associations across two sampling seeds were negative (ρ=−0.828\rho=-0.828 to −0.143-0.143; Appendix F.11, Figures 31-33).

Translation scale. For a random 2,000-readout cohort (500 per area), we measured each readout's receptive field approximately from its MEDI's temporal range map: the intensity range at each pixel. These values, together with horizontal and vertical spatial standard deviations, defined an axis-aligned ellipse. We then scaled horizontal and vertical shifts by the corresponding semiaxis. As predicted, V1 had the lowest median retention at all four tested fractions, 0.25, 0.5, 1, and 2. At one semiaxis, medians were 0.119, 0.156, 0.158, and 0.175 in V1, LM, RL, and AL.

We also matched both MEDI spatial scales across areas, retaining 296 readouts per area. V1 remained the lowest median retention in all 100 selections, both at the original 48-pixel shift and at one semiaxis. The V1 difference therefore persists after controlling the receptive field effects (Appendix F.12, Figures 34-35).

Rotation geometry. For a random 2,000-readout cohort (500 per area), we measured elongation and the image change caused by rotation. Image change was measured by the squared pixel difference divided by the original video's squared deviations from frame means. Elongation had little association with retention (∣ρ∣≤0.043|\rho|\leq0.043); larger image changes were associated with lower retention (ρ=−0.396\rho=-0.396 to −0.315-0.315). Matching both measures still left AL numerically highest, with an unresolved AL-RL difference: 0.0118, 95% interval [−0.0024,0.0252][-0.0024,0.0252], adjusted p=0.262p=0.262. Thus, elongation confounds minimally rotation tolerance, whereas internal pattern changes help interpret moderately (Appendix F.13, Figure 36).

Response offsets. We agree that lifetime sparseness is sensitive to additive offsets. Across the MEDI-readout responses, 44.6% were below the gray-screen response and 48.0% below the black-screen response. Subtracting either screen response and clipping negative values at zero would collapse all below-screen responses to zero, discarding differences among them. Rolls and Tovee (1995) noted this limitation of baseline-subtracted, rectified responses when stimuli suppress firing below the spontaneous rate, and suggested using the lowest stimulus-evoked response as an alternative reference. Following this approach, we recalculated both selectivity and tolerance relative to each readout's minimum response across the shared MEDI set.

For the same random sample of 2,000 readouts (500 per area), we defined mim_i as each readout's minimum response across the 2,000 MEDIs. We subtracted mim_i before computing selectivity and from both responses in the tolerance ratio, [ri(transformed)−mi]/[ri(original)−mi][r_i(\mathrm{transformed})-m_i]/[r_i(\mathrm{original})-m_i]. Within-area preference, highest V1 selectivity, and all four principal tolerance directions persisted. All 24 within-area selectivity-tolerance associations remained negative (ρ=−0.850\rho=-0.850 to −0.189-0.189; all Benjamini-Hochberg-adjusted p<0.001p<0.001). Thus, these findings persist when both metrics use this alternative response reference (Appendix F.10, Figure 30).

AuthorsOpenReview ↗

Response to Reviewer UbuW #3

Critical Weakness 5 — Random versus quality-selected samples

The 200 readouts per area were sampled randomly for baseline comparisons to cover prediction qualities across the released population. The principal area analyses use all 104,171 readouts. We now also test better-predicted readouts.

Within the original baseline cohort (800 readouts), we selected each area's best-predicted quarter by ccnormcc_{\mathrm{norm}} among scored readouts: 47, 48, 47, and 48 targets in V1, LM, RL, and AL. All 20 area-by-baseline median ratios exceeded one (range 1.73-12.20; Appendix C.2, Table 1).

A separate 2,000-readout cohort selected by ccmax⁡cc_{\max}, which was tested in previous experiments in the three new models (2 seeds, 1 transformer head), also retained higher response compared to baseline (Appendix F.2, Figure 21). Among the two experiments, the latter selects readouts by recording reliability, while the former selects readouts by prediction accuracy.

On the form 1-2 — Terminology and hedging

We have manually adapted the writing carefully and substantially, to ensure that all readers stay onboard.

On the form 3 — Interpretation and narrative

The Discussion now links the controls to model sensitivity, temporal contribution, spatial geometry, and stimulus-set dependence. It ends with the biological tests needed.

On the form 4 — Behavior and speed interpolation

Methods 3.1 now states that eye position enters the perspective module, while locomotion and pupil dilation enter the modulation module. All are fixed to zero in normalized model coordinates during synthesis and evaluation.

Methods 3.3 now states that playback changes use linear interpolation along the complete 60-frame sequence, with the output sequence length adjusted to the requested playback speed.

On the form 5 — Known speed preferences

The Discussion links the relative V1 slow-down/speed-up profile to higher temporal-frequency and speed preferences in AL and RL reported by Marshel et al. (2011) and Andermann et al. (2011). We present this as recovery of established tuning differences through the new measure.

On the form 6 — Figures and association shape

The revised panels separate areas by position and color and show medians and interquartile ranges. A dashed line at one marks equal responses; triangles at 20 mark larger ratios. Statistics use uncapped ratios (Figures 2b and 8).

Figures 37-38 show response density and median retention in ten equal-count selectivity groups. Spatial curves approach a low-retention floor; temporal conditions have higher retention and different slopes (Section 4.4; Appendix F.14).

On the form 7 — Unused caption symbol

We removed the unused “n.s.” definition from Figure 3b's caption.

Questions for Authors

Q1, readout confounds: Center-distance matching, scale-normalized translation, rotation geometry, and response-offset controls are detailed under Critical Weakness 4 (Appendices F.10-F.13).

Q2, distribution shape: See On the form 6 and Appendix F.14, Figures 37-38.

Q3, seeds and animals: See Critical Weakness 1 and Appendices F.2-F.3.

Q4, random sampling: See Critical Weakness 5 and Appendix C.2, Table 1.

Q5, writing: See On the form 1-3.

Q6, temporal stationarity: The temporal ablations quantify how static repetition and shuffling affect response. See Critical Weakness 3 and Appendices E.1 and F.6.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank all four reviewers for their careful reading and constructive suggestions. We added tests across three separately trained models, a second mouse, natural-movie comparisons, temporal ablations, and spatial controls. After revision, the paper now reports four main findings:

  1. MEDIs elicit larger predicted responses than all five tested baseline families: dynamic Gabors, static Gabors, most exciting static inputs (MESIs), drifting gratings, and natural movies. We compared each readout's response to its MEDI with its response to each baseline. Same MEDIs consistently elicited larger predicted responses than all stimuli baselines, when they were evaluated by three separately trained response models and by an additional model of a second mouse (Appendices F.2-F.3).

  2. The shared MEDI set reveals within-area preference and the highest selectivity in V1. Readouts in each visual area responded preferentially to MEDIs optimized for readouts in that same area. Selectivity was highest in V1; the ordering of LM, RL, and AL varied across analyses. Within-area preference and highest V1 selectivity persisted across the additional models, optimization and background-noise conditions (Appendices F.1-F.2).

  3. Tolerance follows a different area profile for each transformation. At the selected settings, V1 had the lowest median response retention under translation and speed-up, and the highest median under slow-down; AL had the highest median under rotation. All four median directions persisted in the three additional response models. No single area ordering held across transformations. (Appendix F.2).

  4. Selectivity and tolerance are predominantly negatively associated within areas at the selected settings. The strength of the relationship varies with transformation and parameter. Associations were negative for all six transformations in the original model and across the optimization controls, while speed-up associations were weaker and less consistent across the additional models. (Appendices F.1-F.2).

We have updated the main text, appendices, and accompanying code repository to reflect these changes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.