acceptodds
Under review as a conference paper at ICLR 2027

Learning from Natural Images Untangles Number Sense

Abstract

Numerical cognition is a simple yet fundamental ability that remains challenging for modern vision-language models. Number sense is already present in numerically naïve animals, suggesting that it can emerge without explicit numerical learning. Deep neural networks provide a related observation, with numerosity-selective units appearing after training on unrelated natural-image tasks and even before any training. However, these findings leave unresolved whether visual learning contributes beyond the spontaneous emergence of numerical selectivity. Human development provides a clue, as children become better at making judgments in which numerosity and proportion conflict. Here, we introduce a geometric framework to trace how representations of these two quantities change across learning. Although both untrained and pretrained networks exhibited unit-level numerosity selectivity and similar basic comparison performance, their population-level organization differed substantially. Numerosity and proportion representations overlapped strongly before training but became separated after natural-image learning, supporting more robust readout under conflict. Importantly, this reorganization did not require supervised category learning, as self-supervised learning produced comparable separation while shuffled-label training did not. Together, our results suggest that rudimentary number sense may emerge spontaneously, while structured visual learning reorganizes latent quantity information into representations that enable more independent and robust decoding.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
number sense, numerical cognition, representation learning, representational geometry

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_BEthOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether visual learning improves number sense beyond the emergence of numerosity-selective units. Using dot-array stimuli, it analyzes the population-level geometry of numerosity and proportion representations across network training conditions. The authors report that structured visual learning separates these representations and relate this separation to improved decoding, conflict resolution, cross-condition generalization, and robustness to cross-quantity ablation.

Critical strengths
  1. The paper addresses an interesting, well-motivated question and moves beyond single-unit selectivity toward population-level organization and functional readouts. The behavioral motivation provides a useful perspective on what visual learning might contribute to numerical representations.

  2. The experimental coverage is substantial, including supervised and self-supervised learning, shuffled-label controls, and several synthetic-image distributions. Replication across multiple initializations, together with supplementary ResNet-50 geometry results, strengthens the descriptive findings.

Critical weaknesses
  1. The angle–ablation relationship is sensitive to unexplained exclusions. Figure 6d excludes the noise-trained and label-shuffled networks. Recomputing the correlation from the rounded means in Tables 1 and 3 reproduces approximately (r=-0.88) for the selected states, but only (r=-0.30) when all ten states are included. Low baseline decoding performance could motivate excluding these conditions, but an explicit criterion and sensitivity analysis accounting for baseline decodability are needed before interpreting this relationship as evidence of functional independence.

  2. The connection between the motivating conflict task and the geometric analysis needs clarification. Geometry is defined using total numerosity, (N=W+B), whereas the hard proportion task defines congruency using white-dot count (W) and proportion (P). Figure 6b instead evaluates restricted total-numerosity comparisons under (N/P) conflict (Appendix A.6; Figure S2). These are distinct tests, so the reported correlation does not directly explain the behavioral effect used to motivate the paper. Matched geometry–behavior analyses, explicit restriction thresholds, and joint consideration of congruent and incongruent performance would strengthen the interpretation.

  3. The ablation lacks controls for nonspecific perturbation effects. Appendix A.7 explicitly removes the opposite quantity’s subspace from raw activations without recentering, while keeping the affine decoder fixed. This removes the projected activation mean as well as quantity-dependent variation, potentially shifting decoder outputs independently of information loss. A mean-preserving or analytically bias-corrected ablation, together with matched-rank random-subspace controls, is needed to distinguish representational interference from sensitivity to the intervention.

Questions for authors

See weakness

4 comments
AuthorsOpenReview ↗

Inclusion of all training states, conflict-task clarification, and corrected ablation with direct manipulation.

[Weakness 1] The angle–ablation relationship is sensitive to unexplained exclusions. Figure 6d excludes the noise-trained and label-shuffled networks. Recomputing the correlation from the rounded means in Tables 1 and 3 reproduces approximately (r=-0.88) for the selected states, but only (r=-0.30) when all ten states are included. [...]

- We included all ten training states in the revised angle–ablation analysis and confirmed that the correlation remained significant (Figures 7d, S9; lines 452–454; see Q3 for details of the revised ablation method).

We thank the reviewer for identifying the unexplained exclusions. Following the concern raised in Weakness 3, we replaced the original ablation protocol with a corrected mean-preserving method to avoid shifts in activation means. Using this revised protocol, we repeated the angle–ablation analysis across all ten training states, including Noise-trained and Label-shuffled. The correlation was significant (r = -0.86, p = 0.0014; Figure 7d), providing a more appropriate assessment of the relationship between subspace orientation and cross-quantity interference.

The two states were originally excluded because their baseline numerosity decoding was near chance (Noise-trained: 5.9%; Label-shuffled: 4.9%; chance: 3.3%), substantially below the other conditions (18.7–29.2%). Their baseline errors were also no better than those of a constant-median predictor (Figure S9d,e). We were concerned that such poor baseline decodability would make ablation damage difficult to interpret. However, we acknowledge that this was not an objective exclusion criterion. The revised analysis therefore includes all ten states, and we explicitly report their poor baseline decodability while interpreting the correlation as supporting rather than causal evidence. X\phantom{X} [Weakness 2] The connection between the motivating conflict task and the geometric analysis needs clarification. Geometry is defined using total numerosity, (N=W+B), whereas the hard proportion task defines congruency using white-dot count (W) and proportion (P). [...]

- We clarified the relationship between NW/PN_W/P and N/PN/P conflict and confirmed pretraining-induced separation under both, while focusing subsequent analyses on the more general N/PN/P framework (Figures 2d, S3; lines 223–235).

Thank you for raising this important point. We clarified that the motivating task defines conflict between white-dot count (NWN_W) and proportion (PP), whereas our principal geometric analysis uses total numerosity (NN) and PP. Since NW=NPN_W = NP, NW/PN_W/P-incongruent pairs are a subset of N/PN/P-incongruent pairs, but the two tests are not equivalent (Figure S3a).

We now compare pretraining effects under both conflict definitions (Figure 2d) and report the corresponding NW/PN_W/P and N/PN/P geometries (Figure S3b). Both conflict tasks show a qualitative pretraining advantage, and both geometric analyses show greater separation after pretraining. The task definitions and pair-sampling procedures are described in Appendix A.6, with congruent and incongruent performance reported in Table 1. Together, these analyses clarify the relationship between the motivating NW/PN_W/P conflict task and the broader N/PN/P representational framework, while showing consistent pretraining effects under both definitions. X\phantom{X} [Weakness 3] The ablation lacks controls for nonspecific perturbation effects. Appendix A.7 explicitly removes the opposite quantity’s subspace from raw activations without recentering, while keeping the affine decoder fixed. [...] A mean-preserving or analytically bias-corrected ablation, together with matched-rank random-subspace controls, is needed to distinguish representational interference from sensitivity to the intervention.

- We corrected the ablation protocol to preserve activation means and added matched-rank random-subspace controls (Figures 4, S6, S7, line 149-157)

Thank you for the thoughtful comments. We agree that removing raw projected activity could shift decoder outputs by also removing the projected activation mean. We therefore replaced this procedure with mean-preserving ablation, which removes stimulus-dependent variation along the opposite quantity subspace while preserving the activation mean. We held the decoder fixed and included matched-rank random-subspace ablations as controls (Appendix A.8; Figures S6, S7).

Using this corrected protocol, we further tested functional relevance by manipulating principal angle and representational overlap in pretrained networks across 12 architectures (Figure 4; Appendix A.7). Increasing the angle reduced cross-quantity ablation damage with little effect on decoding accuracy, whereas increasing overlap reduced accuracy and increased damage. Together, these analyses address the methodological concerns about ablation while providing direct functional evidence for the role of representational geometry in quantity readout.

Thanks for the reply! I appreciate the explanation from the authors. However, I think my essential concerns cannot be fully addressed by the reply. Perhaps some careful polishing and rewriting would make this work more ready for publication.

Best,

Reviewer BEth

Thank you for your constructive feedback. We acknowledge that our previous response did not fully address your central concerns. In this further revision, we have made every effort to address these points more directly and thoroughly.

In particular, we have added and substantially revised figures and explanations to clarify:

(1) the explicit exclusion criterion requiring pre-ablation decoding performance to exceed that of a constant predictor (Q1),

(2) the distinction between the NW/PN_W/P and N/PN/P conflict conditions (Q2), and

(3) the importance of mean preservation in the ablation analysis (Q3).

We believe these clarifications will also be valuable to other researchers working on related problems, and we sincerely appreciate your feedback, which has helped us improve the clarity and rigor of our work.

X\phantom{X}

[Q1] The angle–ablation relationship is sensitive to unexplained exclusions. Figure 6d excludes the noise-trained and label-shuffled networks. Recomputing the correlation from the rounded means in Tables 1 and 3 reproduces approximately (r=-0.88) for the selected states, but only (r=-0.30) when all ten states are included. Low baseline decoding performance could motivate excluding these conditions, but an explicit criterion and sensitivity analysis accounting for baseline decodability are needed before interpreting this relationship as evidence of functional independence.

Answer

We agree that our original exclusion criteria were not sufficiently justified. As the reviewer pointed out, assessing ablation-induced damage requires sufficiently high baseline decoding performance for a meaningful decrease to be observed. Based on this rationale, we excluded the noise-trained and label-shuffled conditions because their pre-ablation errors (MAEs) were substantially higher than those of the other conditions.

In the revised manuscript, we have established an explicit exclusion criterion based on a constant predictor that always outputs a fixed value. Specifically, we exclude conditions whose pre-ablation MAE exceeds that of the constant predictor, as ablation-induced damage cannot be meaningfully interpreted when the original decoder already performs worse than this baseline. Under this criterion, the noise-trained and label-shuffled conditions are excluded (Fig. S3, line 902-908).

Finally, to improve transparency and avoid confusion, we have substantially expanded the tables 1-4 to report the numerical values underlying all figures, allowing readers to examine the results directly.

X\phantom{X}

[Q3] The ablation lacks controls for nonspecific perturbation effects. Appendix A.7 explicitly removes the opposite quantity’s subspace from raw activations without recentering, while keeping the affine decoder fixed. This removes the projected activation mean as well as quantity-dependent variation, potentially shifting decoder outputs independently of information loss. A mean-preserving or analytically bias-corrected ablation, together with matched-rank random-subspace controls, is needed to distinguish representational interference from sensitivity to the intervention.

Answer

We acknowledge that our original ablation analysis overlooked the importance of preserving the activation mean, and we sincerely appreciate your valuable feedback.

In the revised manuscript, we introduced mean-preserving ablation to avoid unintended shifts in the activation mean and added matched-rank random-subspace controls to distinguish the effects of removing quantity-related information from nonspecific perturbation effects (Fig. S3 and Appendix A.8).

Furthermore, we have substantially expanded Fig. S3 and Appendix A.8 to explain the differences between ablation methods, their potential confounds, and the importance of appropriate controls. We hope these clarifications will also benefit researchers in cognitive neuroscience who employ similar representational ablation analyses.

[Q2] The connection between the motivating conflict task and the geometric analysis needs clarification. Geometry is defined using total numerosity, (N=W+B), whereas the hard proportion task defines congruency using white-dot count (W) and proportion (P). Figure 6b instead evaluates restricted total-numerosity comparisons under (N/P) conflict (Appendix A.6; Figure S2). These are distinct tests, so the reported correlation does not directly explain the behavioral effect used to motivate the paper. Matched geometry–behavior analyses, explicit restriction thresholds, and joint consideration of congruent and incongruent performance would strengthen the interpretation.

Answer

We agree that our original manuscript did not clearly distinguish the conflict conditions, which may have caused confusion about the connection between the behavioral and geometric analyses.

In the original manuscript, we used NW/PN_W/P conflict in Fig. 2d to align with the experimental paradigm of Hamamouche and Cordes (2023; Fig. 2c), whereas subsequent analyses used N/PN/P conflict. We acknowledge that this distinction was not sufficiently explained.

In the revised manuscript, we now present both NW/PN_W/P and N/PN/P conflict conditions in Fig. 2d, showing that they yield similar trends. We consistently use N/PN/P conflict in subsequent analyses because it provides a more general criterion based on total numerosity (N=NW+NBN=N_W+N_B), directly matching the quantity used in our geometric analyses. Specifically, N/PN/P conflict occurs when the image with greater total numerosity has a lower proportion of white dots (P=NW/NP=N_W/N), or vice versa. We have added Fig. S7 to illustrate these definitions and their relationships. X\phantom{X} We have also directly addressed the three specific points raised in your review:

(1) Matched geometry–behavior analyses:

We have expanded our analyses to include comparisons across architectures as well as training conditions (Fig. 7). We now examine how representational geometry relates to two main functional measures: N/PN/P conflict hard-task performance (Fig. 2d), and ablation-induced damage. These expanded analyses allow us to evaluate the geometry–behavior relationships more consistently across different models and training conditions.

(2) Explicit restriction thresholds:

We have clarified the definition of the hard task, consistently applied across the relevant analyses. Numerosity (NN) takes discrete values from 1 to 30, with a step size of 1, while proportion (PP) has a step size of 1/61/6 (0.167). The hard task compares image pairs differing by exactly one step in both NN and PP. In contrast, the other comparison tasks randomly sample image pairs from the full range of NW,NB∈[1,30]N_W,N_B\in[1,30] (line 236).

(3) Joint consideration of congruent and incongruent performance:

We have clarified why these conditions must be examined separately. Although pretrained models achieve higher overall accuracy than untrained models, both perform well on congruent pairs (Fig. S6). The advantage of pretraining becomes apparent primarily on conflict pairs, where numerosity and proportion favor opposite answers. Thus, overall accuracy can obscure the benefit of pretraining, whereas conflict performance more directly reveals differences in the ability to make correct judgments when competing quantity cues disagree. We have emphasized this distinction in the revised manuscript (Fig. S6).

Furthermore, we have standardized the terminology throughout the manuscript: PP consistently denotes the white-dot proportion (PWP_W) (line 126, 225), and we consistently use "conflict" after initially defining it as equivalent to "incongruent" (line 228).

Although space limitations prevent us from plotting every combination of the expanded analyses, we have substantially expanded Tables 1–4 to report all numerical results, allowing readers to examine these relationships directly.

We sincerely appreciate your feedback, which helped us clarify the experimental conditions and strengthen the connection between our behavioral and geometric analyses.

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

This paper investigates how visual learning shapes numerical representations in neural networks. It shows that natural-image pretraining separates numerosity and proportion representations, improving quantity decoding under conflicting cues. Experiments across supervised, self-supervised, and synthetic training settings support this geometric account.

Critical strengths
  1. The paper moves beyond single-unit numerosity selectivity to population-level representational geometry, providing insight into how visual learning refines number sense.
  2. The study examines multiple training regimes and connects representational separation to decoding performance, generalization, and functional independence.
Critical weaknesses
  1. The relationship between representational separation and improved decoding is primarily correlational. The experiments do not directly establish that geometric separation causes better quantity discrimination.
  2. Most analyses focus on AlexNet, with limited cross-architecture validation. Moreover, some results do not consistently support the proposed relationship between separation and conflict robustness.
Questions for authors
  1. How do the authors explain why the Noise-trained model achieves higher incongruent comparison accuracy than the pretrained model despite substantially greater representational overlap (Tables 1-2)?
  2. Can the authors provide stronger evidence that representational separation directly contributes to improved decoding rather than merely correlating with it?
  3. Can the authors clarify the cross-quantity ablation protocol and extend functional validation to additional architectures?
1 comment
AuthorsOpenReview ↗

Direct functional validation, Cross-architecture generalization, ablation controls, and Noise-trained exception.

[W1] The relationship between representational separation and improved decoding is primarily correlational. [...]

- We moved beyond correlation by directly manipulating representational geometry. (Figs. 4, S7, line 324)

We agree that cross-state correlations alone do not establish functional causality. We therefore directly transformed activations within pretrained networks to manipulate the principal angle between numerosity (N) and proportion (P) subspaces or their representational overlap, approximately preserving the other property (Appendix A.7). Increasing the angle reduced cross-quantity ablation damage with little effect on decoding accuracy (Fig. 4c), whereas increasing overlap reduced decoding accuracy and increased damage (Fig. 4d). These patterns were evaluated across 12 architectures (Fig. S7). Thus, geometry has a functional role in quantity readout, although our interventions do not establish it as the sole mechanism behind pretraining-related improvements in conflict judgments. X\phantom{X} [W2] Most analyses focus on AlexNet, with limited cross-architecture validation. Moreover, some results do not consistently support the proposed relationship between separation and conflict robustness.

- Generalization across architectures (Fig 3c, line 252); exception discussed in [Q1] (Fig. S9)

We expanded the analysis to 10 additional modern architectures, spanning CNNs, Vision Transformers, a VLM visual encoder, and MLP-, state-space-, and MetaFormer-based models. Pretraining consistently increased N–P principal angles and reduced overlap across architectures (Fig. 3c). The functional effects of geometry manipulation were also examined across 12 architectures (Figs. 4, S7).

We agree that geometry does not fully predict conflict accuracy in every training condition, notably the Noise-trained model. We now distinguish this limitation from the reproducible geometric changes and direct functional interventions (Fig. S9; see [Q1]). X\phantom{X} [Q1] How do the authors explain why the Noise-trained model achieves higher incongruent comparison accuracy than the pretrained model despite substantially greater representational overlap (Tables 1-2)?

- Investigating the Noise-trained exception (Tables 1 and 3; Fig. S9, line 456)

The Noise-trained model achieved higher incongruent accuracy than the pretrained model (72.5% vs. 60.6%), despite much greater representational overlap (0.547 vs. 0.083). However, its congruent accuracy was far lower (52.3% vs. 78.9%), as was overall accuracy (56.4% vs. 75.2%). Further analysis showed that its baseline decoding errors, like those of the Label-shuffled model, were no better than those of a constant-median predictor (Figs.S9d, e). We also found strongly imbalanced responses to white- and black-dot counts in these networks (Figs. S9f, g), suggesting atypical sensitivity to different quantity cues.

Together, these observations indicate that the higher incongruent accuracy does not reflect generally better quantity readout. Such response imbalances may contribute to this unusual behavior, although the precise mechanism remains unclear. We therefore acknowledge that representational geometry alone cannot fully predict discrimination performance across heterogeneous training states. X\phantom{X} [Q2] Can the authors provide stronger evidence that representational separation directly contributes to improved decoding rather than merely correlating with it?

- We directly validated the functional role through representational manipulationn (Figs. 4, S7)

As detailed in [W1], we independently manipulated principal angle and overlap within pretrained network representations. Increasing the angle reduced cross-quantity interference without substantially changing decoding accuracy; increasing overlap impaired both decoding and resistance to ablation. These intervention-based results go beyond correlations between training states, demonstrating that geometric organization affects quantity readout. X\phantom{X} [Q3] Can the authors clarify the cross-quantity ablation protocol and extend functional validation to additional architectures?

- We clarified ablation protocol (Fig. S6) and extended functional validation across architectures (Figs. 3, 4, S7)

We clarified the cross-quantity ablation protocol (Appendix A.8). We remove variation along the opposite quantity's subspace while preserving the activation mean, and evaluate damage with the original decoder held fixed. We also included matched-rank random-subspace ablation controls to assess nonspecific perturbation effects (Figs. S6, S7).

Beyond CNNs, we evaluated 10 additional modern models. All showed pretraining-induced geometric separation (Fig. 3c), and direct interventions across the 12 architectures reproduced the expected relationship between larger principal angles and lower ablation damage and between greater overlap and poorer decoding (Figs. 4, S7).

Reviewer_888iOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper examines the role of natural-image training in numerosity and proportion judgments in deep neural networks. The key contribution is to bring neural population geometry analysis (e.g., principal angles and representational overlap) to this question, showing that although untrained and pretrained networks both show unit-level numerosity selectivity and similar basic comparison performance, natural-image training (supervised or self-supervised, but not shuffled-label training) promotes more orthogonal, separated representations of proportion and numerosity, which support more robust readout when the two quantities conflict.

Critical strengths
  1. Systematic analysis of population-level geometry. Rather than relying only on unit-level numerosity selectivity, the paper systematically characterizes population-level representational geometry (principal angles, representational overlap) and shows that natural-image training separates numerosity and proportion representations. This moves the field beyond the observation that number-selective units emerge spontaneously and offers a concrete account of what visual learning adds.

  2. Clarity and grounding. The paper is well written and easy to follow, with a strong background in the related cognitive science and neuroscience literature.

Critical weaknesses
  1. Limited model scope and scale. The paper relies on older models such as AlexNet and considers only convolutional neural networks, not other architectures such as transformers. The experiments are also small in scale, so it remains open whether the findings generalize to modern, scaled-up models.

  2. Impoverished stimuli and uncontrolled confounds. The visual stimuli are very impoverished, and potential confounds from other visual features are not well controlled. In particular, the comparison between untrained and pretrained networks is confounded by stimulus diversity: the pretrained model has experienced a broad range of naturalistic visual input, which may by itself contribute to its better performance on the proportion task, where black and white dots are shown simultaneously. Without controlling for this, the improved proportion/numerosity separation cannot be attributed specifically to structured visual learning.

Questions for authors
  1. (Control analysis) To address the stimulus-diversity confound, could you examine the numerosity comparison task using stimuli that contain both black and white dots (matching the proportion-task displays)? If the pretrained advantage reflects exposure to mixed-contrast stimuli rather than a reorganization of quantity representations, the untrained vs. pretrained difference should change on this task. This could materially affect my assessment of the main claim.

  2. (Clarification/generalization) Do the geometric results hold for non-convolutional architectures (e.g., vision transformers) and for larger, more modern models?

  3. (Clarification) How were low-level visual features (beyond those already controlled) handled in the stimulus design, and do the main results survive when those features are further controlled?

3 comments
AuthorsOpenReview ↗

Cross-architecture generalization and additional visual controls.

[W1] Limited model scope and scale. [...]

- We expanded our analysis to 12 diverse architectures and confirmed consistent pretraining-induced quantity separation (Figure 3c, line 252-257)

We thank the reviewer for highlighting the limited model scope. AlexNet was originally chosen to compare directly with previous number-sense studies. We have now added 10 modern models spanning CNNs (ResNeXt-50, EfficientNet-B0, DenseNet-121), Vision Transformers (ViT-B/16, Swin-T, DeiT III-L/16), a VLM encoder (UniCL Swin-T), and alternative architectures (gMLP-S, Vim-S, PoolFormer-S24). Across these diverse architectures, supervised pretraining consistently increased N–P principal angles and reduced representational overlap (Figure 3c). We also observed similar separation across self-supervised models (Figure 5). Thus, the geometric reorganization is not specific to convolutional architectures. X\phantom{X} [W2] Impoverished stimuli and uncontrolled confounds. [...]

- We expanded stimulus diversity and confirmed robust quantity separation across additional low-level visual controls (Figures 2b, S2, S5, line 179-185)

We expanded the visual controls to address this concern (Q1 and Q3). Untrained networks reliably discriminated numerosity even with mixed-color dots, and pretraining yielded a substantially larger advantage on incongruent than congruent quantity comparisons. Additional low-level visual controls confirmed that representational separation persisted across distinct stimulus regimes. Importantly, supervised and shuffled-label networks received the same ImageNet images but exhibited different geometric organization, whereas label-free self-supervised learning also produced separation. This suggests that exposure to diverse images alone is insufficient; how the network learns from visual structure matters. We do not claim that these controls completely isolate the underlying learning mechanism. X\phantom{X} [Q1] (Control analysis) Could you examine the numerosity comparison task using stimuli that contain both black and white dots? [...]

- We added mixed-color numerosity controls and confirmed that untrained performance remains high (Figures 2b, 2d, S2, line 179-185)

We extended the numerosity comparison to eight dot-color and background conditions, including mixed black-and-white stimuli matching the proportion task. Untrained networks maintained high accuracy (79.5–84.7%) across conditions, while pretrained networks showed modest improvements of 0.9–6.4 percentage points (Figure S2). Under the same gray-background conditions, the pretrained advantage was comparable for single-color and mixed-color dots, suggesting that the pretrained advantage was not specific to mixed-color processing. Moreover, in the same mixed-color proportion task, pretraining improved accuracy by only 2.0 percentage points on congruent trials (76.9% to 78.9%), compared with 10.7 points on incongruent trials (49.9% to 60.6%; Figure 2d). Although task difficulty may contribute, the larger improvement under conflict is unlikely to reflect merely better processing of mixed-color stimuli. We also clarified that single-color dots were used only for the initial comparisons with previous studies (Figures 2a,b); subsequent experiments used mixed black-and-white dots. X\phantom{X} [Q2] (Clarification/generalization)

- We expanded our analysis to 12 diverse architectures and confirmed consistent pretraining-induced quantity separation (Figure 3c, line 252-257)

As detailed under Weakness 1, the geometric results extend to 10 additional modern models, including Vision Transformers, MLP-based models, and state-space models. All showed increased principal angles and reduced overlap after supervised pretraining. X\phantom{X} [Q3] (Clarification) How were low-level visual features handled in the stimulus design? [...]

- We added low-level visual feature controls and confirmed robust quantity separation across different stimulus conditions (Figure S5, line 255-257)

The original stimuli comprised three sets separately controlling dot size, total area, and spatial extent. We analyzed geometry and decoding independently within each set rather than only pooling them. Pretraining consistently increased principal angles, reduced overlap, and improved N and P decoding across all three sets (Figure S5c). Thus, the effects do not arise simply from combining stimuli with different low-level statistics. We further tested whether area cues explained the decoding advantage. Using fixed readouts, we evaluated generalization when area-based cues were aligned with, balanced against, or reversed relative to count-based quantities. Pretrained networks retained an advantage for numerosity and proportion under these conditions (Figures S5d,e). Together, the results support robustness to the examined low-level properties without claiming to exclude every possible visual confound.

Thank you for the thorough response and revision. The extension to 12 architectures addresses my concern about model scope, and the mixed-color control largely rules out the stimulus-diversity account I raised. However, I remain uncertain about the effect size, and I am still concerned that the findings are limited to simple synthetic dot-array stimuli. Therefore, I will keep my score.

Thank you for taking the time to review our revisions and for your constructive feedback. We agree that extending the analysis beyond synthetic dot arrays would be an interesting direction, and we wish we could have explored it in this revision. We sincerely appreciate your thoughtful comments.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Summary of Major Revisions

We sincerely thank the reviewers for their constructive feedback, which led us to substantially revise both the analyses and the overall structure of the manuscript.

The revised manuscript now centers on two major advances:

1. Cross-architecture generalization: We demonstrated consistent pretraining-induced quantity separation across 12 diverse architectures (Figure 3; addressing Reviewers 888i and wdpR).

2. Direct functional validation: We validated the functional role of representational geometry through direct manipulation across these architectures (Figure 4; addressing Reviewer wdpR).

Whereas the previous version focused on comparing various training states within a limited set of CNNs, we substantially reorganized the manuscript to emphasize the generalizability and functional consequences of pretraining-induced quantity separation (Figures 3 and 4).

Analyses of different training conditions now serve as supporting evidence that neither category labels nor natural-image exposure is strictly necessary for quantity separation, as demonstrated by self-supervised learning and synthetic-image training, respectively (Figures 5 and 6).

We also corrected the cross-quantity subspace ablation protocol using mean-preserving ablation, and clarified the relationship between NW/PN_W/P and N/PN/P conflict by examining both definitions and focusing subsequent analyses on the more general N/PN/P framework (addressing Reviewer BEth and wdpR).

We believe these revisions provide a clearer and more compelling account of how visual learning refines number-sense representations.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.