acceptodds
Under review as a conference paper at ICLR 2027

ReGraph: A Computational Account of Emergent Generalization in the "What" and "Where" Dual Visual Streams

Abstract

Where generalization capacity—the ability to extract context-invariant relational structures—first emerges remains a central question in AI and neuroscience. In the brain, the foundation for this capacity lies upstream of the hippocampus, within the entorhinal cortex, where parallel pathways dissociate relational structure in the medial entorhinal cortex (MEC) from sensory content in the lateral entorhinal cortex (LEC). However, as Eichenbaum argued, this structure–content factorization likely originates earlier, driven by the segregation of the dorsal ('where') and ventral ('what') visual streams. Supporting this, grid-like firing patterns—a defining cellular signature of MEC context-invariant representations—also appear in preceding neocortical regions (e.g., retrosplenial cortex) along the dorsal pathway. While these findings suggest that relational structures do not arise *de novo* in the hippocampal formation, how such representations are computationally formed along upstream pathways remains unknown. To investigate this *in silico*, we developed **ReGraph**, a dual visual-stream model implemented on a recurrent graph architecture with several biological inductive biases, including retina-driven stream-specialized encoding, dorsal-to-ventral modulation, and dynamic lateral connectivity. Trained on the action benchmark Something-Something V2, ReGraph revealed a pathway-specific emergence of relational mapping: context-invariant codes and grid-like spatial bases uniquely co-emerged along the extended dorsal stream. In contrast, the absence of these representations in single-stream models, unmodulated dual-stream variants, and standard action-recognition baselines (e.g., VideoMAE, SlowFast) implies that our selected inductive biases are indeed critical prerequisites for the emergence of relational structures. Crucially, the grid-like bases turned out not to be mere byproducts of architectural design, as our post-hoc analyses demonstrated that these bases could serve as reusable routing templates for information processing via lateral connectivity. Together, our findings provide a computational account suggesting that generalization may not be a faculty that emerges abruptly within a dedicated region, but a property that already takes shape as sensory information is parsed into factorized streams of hierarchical visual processing.

Primary area:
applications to neuroscience & cognitive science
Keywords:
dual visual streams, relational structure, grid-code, biologically inspired neural networks

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_dpx4OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

The authors investigate whether representations supporting relational generalization can develop in a model of the visual pathways upstream of the hippocampal-entorhinal system. ReGraph combines asymmetric dorsal/ventral inputs and task-specific pretrained feature extractors with recurrent self-attention layers and dorsal-to-ventral cross-attention. It is trained on a balanced 33-class Something-Something V2 subset. The proposed Relational Primitive Analysis combines spatial HOSVD of pre-softmax attention scores, gridness tests, object-disjoint linear probing, basis ablation, and reconstruction of unseen attention scores. The principal findings are an increasing fraction of grid-like bases in the full model's dorsal hierarchy, improved late ventral probing performance with dorsal modulation, and greater late-layer sensitivity to high-gridness basis ablation. The authors interpret these findings as evidence for reusable routing templates supporting context-invariant representations.

Critical strengths
  1. The paper makes an interesting empirical connection between visual-stream modeling and grid-like structure in learned attention. Its distinctive contribution is the investigation of spatial connectivity bases in a video action-recognition model, extending earlier work on dual visual streams and structural representations in hippocampal models. The dorsal grid-like fraction rises from 0.0% to 15.3% across four layers, with the qualitative trend retained at 3%, 5%, and 7% singular-direction retention (Tables 3 and 14).

  2. The analysis goes beyond visual resemblance. Appendix L.1 intervenes on attention scores during the forward pass and evaluates the resulting representations with newly fitted OOD probes. At zero added noise, high-gridness ablation produces accuracies 7.0 and 6.0 percentage points below low-gridness ablation at L3 and L4, respectively. Complementary reconstruction results favor high-gridness bases in these layers across the reported 30-50% candidate pools and k=5/9 settings (Table 18). These experiments provide meaningful evidence that the identified structures are functionally relevant.

Critical weaknesses

Appendix L.2 shows that training-derived subspaces capture unseen score structure, with coefficients computed from each unseen matrix. It does not directly establish how periodic geometry supports transfer through normalized attention and transmitted values. The forward-pass ablation provides functional evidence, but reports newly trained probes' OOD accuracy without an ID comparison establishing transfer specificity. Reconstruction and ablation also use different candidate pools and selection rules, with no reported overlap of the selected bases. Thus, the combined analyses do not yet establish reuse and causal relevance of the same templates. The stronger reusable-routing interpretation of the gridness in bases needs additional support or narrower wording.

The claim that the selected biological biases are critical prerequisites for relational representations exceeds the comparisons. Table 2 already shows strong OOD decoding in the unmodulated dorsal stream: 96.9% at L4 versus 97.4% in full ReGraph, despite their different grid-like fractions (6.3% versus 15.3%, Table 3). Thus, strong cross-object action decodability must be distinguished from the particular joint gridness/probing pattern reported in ReGraph. The D-to-V removal is an informative controlled comparison, but the other trait ablations bundle changes, and single-stream references differ in capacity, token budget, and training recipe (Table 1; Appendix H.2). The strongly specialized pretrained front ends further limit attribution of the findings specifically to retinal asymmetry. Appendix M also acknowledges one training seed per configuration. These results support a hypothesis about this model's organization, but not the stronger prerequisite claim or a firm account of where generalization originates.

The architectural positioning is also incomplete. SlowFast is cited, but its M/P-inspired temporal asymmetry and fast-to-slow fusion are not discussed as precedents. The manuscript does not cite the Tolman-Eichenbaum Machine 2020, which links structure/content factorization to generalization across environments, or its transformer formulation in 2022 ICLR, which produces grid-like recurrent positional representations. These are close conceptual precedents, and better connected with the biological observations. These results do not duplicate ReGraph's gated cross-attention, but warrant a more precise account of which design principles and computational explanations are new.

The missing references: TEM 2020: Whittington, J. C. R., et al. (2020). The Tolman-Eichenbaum Machine: Unifying Space and Relational Memory through Generalization in the Hippocampal Formation. Cell, 183, 1249-1263.e23. https://doi.org/10.1016/j.cell.2020.10.024 The transformer variant: Whittington, J. C. R., Warren, J., and Behrens, T. E. J. (2022). Relating transformers to models and neural representations of the hippocampal formation. ICLR. https://openreview.net/forum?id=B8DVo9B1YE0

Questions for authors
  1. Were the probe-test object identities excluded from the 33-class action-recognition training of ReGraph and the associated fine-tuning of its feature extractors, or only from probe training? Please specify how object identities and the train/test sets were constructed. This affects whether the evaluation establishes transfer to objects unseen by the representation learner or by the probe.

  2. How many high-gridness reconstruction bases are also in the actual high-gridness ablation set, separately for source and target modes? I guess the expanded pool may reasonably improve gridness separation, but the concern is the correspondence between the sets supporting the two conclusions.

  3. Please distinguish the contribution from TEM and its transformer formulation (see above). Do the authors claim that these subspaces support generalizable features, or that their periodic geometry implements reusable routing? What evidence distinguishes the latter interpretation? A direct analysis of attention-output changes using a common selected subspace, or an ID-versus-OOD comparison of the existing intervention, would help.

  4. Scope of the prerequisite claim: How should the abstract's claim be reconciled with the strong unmodulated dorsal OOD results? Is the intended claim specifically about the observed combination of increasing gridness and strong decoding within the tested model family?

2 comments

Dear Reviewer dpx4,

Thank you very much for your insightful critique and helpful feedback. In the following responses, we have tried our very best to comprehensively address every point you raised. We hope our clarifications meet your expectations and look forward to any further discussion.

(1) W6: “Single-seed limitation”

We agree with the reviewer that evaluations based on a single training seed may be statistically insufficient to support our strong claim. To address these concerns, we evaluated both Top-1 performance and layer-wise grid-like code ratios across all seven model variants using multiple seeds. Table1 reports the mean results for both metrics (averaged across n = 3 seeds for ReGraph, and n=2 for the others due to computational constraints). We are actively running additional seeds and will provide the updated results as soon as they become available, in addition to incorporating the full statistics into the revised manuscript.

Across all seven variants, ReGraph achieved the highest task performance. Furthermore, an increasing pattern of grid-like code ratios across successive layers of the extended dorsal stream (reaching a maximum of 12.20% at L4, with the highest mean of 5.30) was observed only in the full model, in which all biological inductive biases implemented across the primary and extended streams were combined. These results demonstrated the robustness of our findings for the contribution of the proposed architectural design to both task performance and the hierarchical development of grid-like representations.

[Table1. Top-1 Accuracy and Layer-wise Grid-like Code Ratio across all variants]

VariantStreamTop-1L1L2L3L4Mean L1–L4Monotonic
ReGraphdorsal74.69 (0.17)0.023.195.7912.205.30O
ventral0.186.616.486.845.02X
Dorsal-trait Onlydorsal72.50 (0.11)0.025.555.565.374.12X
ventral0.536.817.095.264.92X
Unmodulateddorsal72.48 (0.32)0.038.413.587.544.89X
ventral0.106.126.755.714.67X
Dorsal-Onlydorsal71.65 (0.06)0.005.297.135.004.35X
Ventral-trait Onlydorsal70.79 (0.51)0.006.608.712.994.57X
ventral0.075.505.625.464.16X
Trait-Symmetricdorsal69.78 (0.01)0.016.414.345.354.03X
ventral0.476.687.375.064.89X
Ventral-Onlyventral60.87 (0.69)0.125.775.455.544.22X

(2) W3 & Q2: “Correspondence between reconstruction and ablation bases for reusable routing.”

We verified that 97% to 99% of the grid-like bases used in Bases-Ablation are indeed included in the Reconstruction's GhighG_{high} set in L3–L4 (L3: Target 98.9% / Source 98.7%, L4: Target 96.9% / Source 98.1%). However, to further completely rule out any residual effects from other bases introduced by expanding the candidate pool, we re-performed the Connectivity-Reconstruction analysis using strictly the exact same grid-like and non-grid-like bases as those used in the Bases-Ablation. As a result, the reconstruction score (R2R^2) of GhighG_{high} (0 → 0 → 0.103 → 0.116) exhibited a distinct surge in the later layers (L3–L4) compared to GlowG_{low} (0 → 0 → 0.013 → 0.073), clearly reproducing the exact same trend as our original conclusion.

(3) Q1: In our original experimental setup, an object-disjoint OOD split was used only for training and evaluating the linear probes, whereas the ReGraph representation learner itself was trained using the standard SSV2 split. We acknowledge that this leaves open the possibility that the representation learner may have indirectly acquired information about the probe-test objects during its training. To address this limitation and more rigorously evaluate genuine context invariance, we conducted additional cross-dataset linear probing experiments using four entirely new motion/action datasets that ReGraph had never encountered during training: KTH Actions, Moving-MNIST, TartanAir, and LIBERO (Table 1). These datasets are entirely different from SSV2 in object appearance, background characteristics, and types of motion. The “D–V (avg)” column in Table 1 represents the performance gap between the average performance of the four dorsal-stream layers and that of the ventral stream for each dataset. As shown in Table1, the dorsal stream consistently outperformed the ventral stream across all four unseen datasets (+3.2%p to +38.1%p). These findings demonstrate that the dorsal stream is particularly specialized in extracting context-invariant spatiotemporal structure, even when evaluated on entirely unseen datasets.

Table 1: Motion / Action Probes (Dorsal > Ventral).

DatasetChance (%)StreamL1L2L3L4AvgD–V (avg)
KTH actions (Schüldt et al., 2004)16.7D81.479.981.579.680.6+8.0
V75.868.675.570.772.6
Moving-MNIST heading (Srivastava et al., 2015)12.5D96.586.485.177.886.5+38.1
V47.041.160.544.848.3
TartanAir rotation sign (Wang et al., 2020)50.0D78.574.177.474.476.1+14.0
V59.959.667.860.962.0
LIBERO z-displacement sign (Liu et al., 2023)50.0D99.899.699.699.599.6+3.2
V95.796.598.095.596.4

(4) W5: "Complementary Stream Asymmetry vs. Capacity and Pretrained Front-End Confounds?" We agree that our ablation studies involve changes in multiple factors, including model capacity and token budget, and therefore should not all be interpreted as strictly one-to-one controlled comparisons. However, these changes are inherent and necessary to the design required to test our hypothesis. Our central question is how the holistic synergy of biologically motivated properties—M/P-driven asymmetry, parallel dorsal/ventral processing, and cross-stream interactions—supports the emergence of relational primitives. Because ablating into a single-stream model naturally reduces both the parameter count and token budget, we adopted a slightly higher learning rate for these smaller models (Appendix H.2); rather than confounding the evaluation, this customized training recipe was specifically necessary to ensure their optimal convergence and provide a genuinely fair comparison. We also agree that the specialized pretrained front-ends make it difficult to attribute the observed effects exclusively to retinal asymmetry. However, we would like to clarify that our proposed M/P-driven asymmetry is an integrated biological inductive bias, incorporating both the complementary spatiotemporal properties of retinal inputs and the functional differentiation of early visual processing. Our hypothesis concerns how these complementary properties jointly shape downstream representations, rather than whether retinal sampling alone is sufficient to produce relational structure.

Nevertheless, to address the reviewer's concern more directly, we conducted additional experiments with two new trait-symmetric dual-stream variants: Dual-Dorsal and Dual-Ventral. In these models, both pathways receive identical retinal-style inputs and employ the same pretrained front-end, thereby removing the complementary specialization between streams while preserving the dual-stream framework. Both variants therefore have parameter capacities and token budgets equal to or greater than those of ReGraph.The results further support our hypothesis: neither variant exhibited the characteristic progressive increase in significant grid-like bases observed along ReGraph's extended dorsal stream (0.0% → 5.1% → 8.8% → 15.3%). In Dual-Ventral, the corresponding proportions were 0.20% → 6.70% → 6.53% → 5.60% in the first stream and 0.29% → 5.69% → 5.50% → 5.53% in the second stream. Similarly, neither stream in Dual-Dorsal exhibited a progressive increase (1st: 0.00 → 4.97 → 4.01 → 5.56 → 3.63, 2nd: 2.45 → 5.23 → 2.18 → 4.56 → 3.61). These results provide critical evidence that complementary M/P-driven dorsal–ventral asymmetry plays a key role in supporting the emergence of relational primitives and the improved performance observed in ReGraph.

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

The paper is motivated by the hypothesis that the factorization of structure and content in the medial and lateral entorhinal cortex (MEC and LEC), for example the separate encoding of an environment’s spatial layout and the objects it contains, originates upstream in the dorsal and ventral visual streams. It asks how structure representations with properties typical of MEC, such as grid-like firing and context invariance, are formed along these upstream pathways. The authors propose ReGraph, a biologically constrained dual-stream model of the dorsal and ventral pathways. In each pathway, a primary visual stage simulates the different properties of magno- and parvocellular retinal inputs and the distinct processing in the two streams, using pretrained models associated with spatiotemporal (dorsal) and object-oriented (ventral) processing. An extended visual stage then models the higher-order cortical hierarchy with layers that have lateral and recurrent connectivity, and a dorsal-to-ventral modulation routes signals from higher dorsal layers to lower ventral layers. The model is trained on action recognition. The authors report that grid-like and context-invariant representations arise progressively along the dorsal stream, and that context invariance also appears in late ventral layers through dorsal-to-ventral modulation.

The paper claims five contributions: a dual-stream model of the primary and extended visual pathways with biological inductive biases; a demonstration of dorsal-to-ventral feedback effects; a framework for quantifying relational primitives layer by layer; evidence that these biases collectively support the progressive emergence of grid-like and context-invariant codes in the dorsal stream; and ablation evidence that high-gridness representations in later dorsal layers support context invariance. From this, the authors conclude that the computational basis for generalization may emerge from the factorization of the dual visual streams rather than exclusively within the hippocampal-entorhinal system. In my reading, the question is exciting and the approach of jointly modeling the dorsal and ventral streams as input pathways to MEC and LEC is novel. However, the model rests on many assumptions, and it remains unclear which findings follow from these assumptions and which emerge from learning. In addition, the conclusion is framed in terms of factorization, but the analyses address mainly the structure side in the dorsal stream.

The authors compare task performance against baselines in the field. The full model falls between them, which is reasonable for a heavily constrained model. Across ablations, performance differences are small, and the Dorsal-Only and Dorsal-trait Only variants already perform close to the full model. The layer-wise analysis is more informative: only the full model shows an increasing proportion of grid-like bases along the dorsal stream, while Dorsal-Only, Unmodulated, and standard baselines do not.

Critical strengths

Exciting question and novel approach: How context invariance and generalization emerge upstream of the hippocampus, and how the dorsal and ventral visual streams shape MEC and LEC representations, is an exciting question. Studying it in silico with a model informed by knowledge of the visual system is a good approach, and modeling both visual and extended pathways jointly is novel. The model yields experimentally testable predictions. The most novel is that context invariance in later ventral stages depends on dorsal input, which could be tested by perturbing dorsal areas. The model also predicts that gridness increases progressively along the dorsal hierarchy, which extends existing reports of grid-like activity outside the entorhinal cortex.

Clear presentation and solid evaluation. The paper is well written, the figures are clear, and the main methods are described at the right level of abstraction. The authors compare against baselines and include ablations.

Critical weaknesses

Unclear what is built in and what emerges: The model incorporates many biologically motivated features: asymmetric magno- and parvocellular inputs, pretrained networks for the two streams, the ReGraph layer design, and dorsal-to-ventral modulation. Most are reasonable choices, but it is not clear which observations follow directly from these design choices and which emerge from learning. For example, the dorsal stream receives low-resolution, grayscale input, which may already make it invariant to appearance-based context changes. Similarly, the dorsal-to-ventral modulation injects dorsal representations into the ventral stream, so a linear probe on late ventral layers can be expected to recover their context invariance. Since a linear probe can exploit even a weak, low-variance signal, decodability does not show that the ventral representation itself has become context-invariant. A control is needed to show that this invariance is not simply inherited during test-time, for example by removing the dorsal contribution at test time, or by projecting out the subspace spanned by the injected dorsal signal before probing. The existing ablations show that removing a component reduces performance or gridness. It would be nice to understand further what leads to the emergence of properties, beyond the ablations performed in the paper.

Conclusions go beyond the quantitative evidence: In task performance, the differences between the full model and the strongest ablations are small (74.57% vs. 71.69% to 72.42%), and the layer-wise gridness analysis covers only two of the ablations. The performance differences are not reported with error margins, for example across initializations, so it is not possible to tell whether the differences are robust. In addition, the paper concludes that the computational basis for generalization emerges from the factorization of the dual visual streams. The analyses, however, address mainly the structure side in the dorsal stream, and they do not test whether the ventral stream carries content separately. The authors should adjust the scope of the paper's motivation / clarify this, since the paper’s conclusions depend on these comparisons and on the factorization framing, and in their current form neither is established.

Questions for authors

Does context invariance in late ventral layers persist when the dorsal contribution is removed at test time, or when the subspace spanned by the injected dorsal signal is projected out before probing?

Which design choices are necessary for the emergence of grid-like bases? Is the gridness analysis also available for the remaining ablations (Trait-Symmetric, Dorsal-trait Only, Ventral-trait Only)?

Are the differences between the full model and the Dorsal-Only, Dorsal-trait Only, and Unmodulated variants significant? What are the error margins across initializations, for both task performance and the proportion of grid-like bases?

Given that the ablated models have fewer parameters, how do the authors rule out that performance differences reflect model capacity rather than the biological constraint?

Section 3.2 includes Dorsal-Only in the layer-wise analysis, but point 1 discusses context invariance only for Unmodulated and ReGraph. How context-invariant is the dorsal stream in Dorsal-Only? Since Dorsal-Only lacks the increasing trend in grid-like bases, this comparison would show whether context invariance depends on them.

The paper concludes that generalization may emerge from the factorization of the dual visual streams, but the analyses focus on the structure side, mainly in the dorsal stream. Does the model show that the ventral stream carries content separately from structure?

9 comments

Dear Reviewer LCSw,

We appreciate your valuable comments, which helped us improve our work. We provide our detailed responses below and hope they adequately address your questions and concerns. We would also welcome any further discussion or requests for clarification.

(1) W2&W5&Q1&Q6. We agree that the high OOD performance of late ventral layers reflects context-invariant relational information inherited from the dorsal stream via dorsal-to-ventral (D→V) modulation, rather than originating within the ventral stream itself. In your suggested control experiment—severing D→V modulation in ReGraph at test time and retraining the linear probes—the ventral OOD accuracy dropped from the intact ReGraph performance across all layers (Table 1; e.g., from 97.2% and 96.2% at L3–L4 to 57.1% and 58.0%), falling to levels lower than those of the unmodulated baseline (Table 1; 64.1% and 69.3% at L3–L4). This raises a further question: how much then does the ventral stream retain its own functional specialization beyond receiving dorsal-derived information? To address this, we examined the contribution of D→V modulation and the distinct functional roles of the two streams. As reported in our manuscript, the unmodulated model provides a direct comparison for evaluating the functional contribution of D→V modulation. Without this modulation, the ventral stream achieves substantially lower OOD accuracy at L3 and L4 (59.2% and 58.4%, respectively), compared with 95.9% and 95.0% in ReGraph. These results indicate that D→V modulation does make an important contribution to the OOD performance of the ventral stream. To further characterize the functional role of the ventral stream beyond this dorsal contribution, we conducted additional layer-wise linear probing analyses on eight external datasets, comprising four motion/action datasets and four object/appearance datasets (Table2/3). We froze the weights of ReGraph pretrained on SSV2 and evaluated representations from successive layers of both streams. The results revealed a clear double dissociation between the two streams. Specifically, the dorsal stream consistently outperformed the ventral stream across all layers on motion/action tasks, with positive D–V differences (Table2, “D–V (avg)” column). Conversely, the ventral stream consistently outperformed the dorsal stream on object/appearance tasks, with negative D–V differences. Together, these findings show that the ventral stream retains its capacity to represent object- and appearance-related information while integrating dorsal-derived relational information that contributes to its OOD action recognition performance (Table 2). Thus, the ventral stream is not merely a passive recipient of dorsal representations; rather, it maintains its own functional specialization while benefiting from information transmitted through D→V modulation. These results support the complementary functional organization of the two interacting streams.

Table 1. Comparison of layer-wise ventral OOD accuracy (%) across ReGraph, severed ReGraph, and Unmodulated.

Ventral LayerReGraph (Intact)ReGraph (D→V Severed)Unmodulated
165.049.156.0
266.753.362.5
397.257.164.1
496.258.069.3

Table 2: Motion / Action Probes (Dorsal > Ventral).

DatasetChance (%)StreamL1L2L3L4AvgD–V (avg)
KTH actions (Schüldt et al., 2004)16.7D81.479.981.579.680.6+8.0
V75.868.675.570.772.6
Moving-MNIST heading (Srivastava et al., 2015)12.5D96.586.485.177.886.5+38.1
V47.041.160.544.848.3
TartanAir rotation sign (Wang et al., 2020)50.0D78.574.177.474.476.1+14.0
V59.959.667.860.962.0
LIBERO z-displacement sign (Liu et al., 2023)50.0D99.899.699.699.599.6+3.2
V95.796.598.095.596.4

Table 3: Object / Appearance Probes (Ventral > Dorsal)

DatasetChance (%)StreamL1L2L3L4AvgD–V (avg)
Flowers-102, 20 of 102 species (Nilsback & Zisserman, 2008)5.0D49.443.345.641.645.0-14.3
V70.056.653.457.259.3
Oxford-IIIT Pets, 20 of 37 breeds (Parkhi et al., 2012)5.0D27.622.123.820.623.5-11.2
V48.530.529.630.234.7
LIBERO-Object target object, 10 tasks (Liu et al., 2023)10.0D64.762.970.866.566.2-22.2
V96.190.483.983.588.5
TartanAir environment, 6 scenes (Wang et al., 2020)16.7D66.452.156.554.557.4-14.6
V84.671.167.464.872.0

(2) W4&Q2&Q3. Following the reviewer's comment, we evaluated both Top-1 performance and layer-wise grid-like code ratios across all seven model variants using multiple seeds. Table1 reports the mean results for both metrics (averaged across n = 3 seeds for ReGraph, and n=2 for the others due to computational constraints). We are currently actively running additional seeds and will provide the updated results as soon as they become available, in addition to incorporating the full statistics into the revised manuscript. Across all 7 variants, ReGraph achieved the highest task performance. Furthermore, an increasing pattern of grid-like code ratios across successive layers of the extended dorsal stream (reaching a maximum of 12.20% at L4, with the highest mean of 5.30) was observed only in the full model, in which all biological inductive biases implemented across the primary and extended streams were combined. These results provide broader quantitative support for the contribution of the proposed architectural design to both task performance and the hierarchical development of grid-like representations.

[Table1. Top-1 Accuracy and Layer-wise Grid-like Code Ratio across all variants]

VariantStreamTop-1L1L2L3L4Mean L1–L4Monotonic
ReGraphdorsal74.69 (0.17)0.023.195.7912.205.30O
ventral0.186.616.486.845.02X
Dorsal-trait Onlydorsal72.50 (0.11)0.025.555.565.374.12X
ventral0.536.817.095.264.92X
Unmodulateddorsal72.48 (0.32)0.038.413.587.544.89X
ventral0.106.126.755.714.67X
Dorsal-Onlydorsal71.65 (0.06)0.005.297.135.004.35X
Ventral-trait Onlydorsal70.79 (0.51)0.006.608.712.994.57X
ventral0.075.505.625.464.16X
Trait-Symmetricdorsal69.78 (0.01)0.016.414.345.354.03X
ventral0.476.687.375.064.89X
Ventral-Onlyventral60.87 (0.69)0.125.775.455.544.22X

(3) Q4. We understand the reviewer's concern that model capacity may account for the performance differences between ReGraph and the ablated models. To examine this possibility, we compared ReGraph with two models with larger parameter counts: an additional Dual-Ventral control model and the external baseline TimeSformer. First, we designed and evaluated a Dual-Ventral control model in which the dorsal stream was replaced with a second ventral pathway, so that both streams received high-resolution RGB inputs processed by the pretrained networks of the primary ventral stream. The Dual-Ventral model contained 96,581,475 parameters, approximately 7% more than ReGraph (90,236,420 parameters), yet achieved a Top-1 accuracy of only 63.06%, 11.51 percentage points below ReGraph (74.57%). Second, we examined TimeSformer as an external baseline (Appendix K). Despite having approximately 121.28 million parameters, it achieved a Top-1 accuracy of 62.3% on our SSV2 subset, 12.27 percentage points below ReGraph. Together, these comparisons suggest that ReGraph’s performance advantage is associated not simply with model capacity, but with the combination of distinct dorsal and ventral processing pathways and their biologically motivated inductive biases.

(4) W1: Unclear what is built in and what emerges.

The reviewer raised an important question. To address the concern that our findings are merely built-in byproducts of dorsal-specific design rather than properties emerging through learning, we provide three pieces of counter-evidence. First, regarding the standalone limitation of dorsal-specific input (Table 3), if the dorsal input pattern alone trivially guaranteed an invariant topology, the same phenomenon would appear in the single-stream Dorsal-Only model without a ventral pathway. However, as shown in Table 3, the Dorsal-Only model exhibits no monotonic layer-wise increase in grid-like bases (0.0% → 5.6% → 4.0% → 2.6%). Second, to address the potential counter-argument that the Dorsal-Only failure stems merely from the limited model capacity of a single-stream structure—or that dorsal input would trivially yield this pattern once combined with the structural capacity of a dual-stream architecture—we conducted a new control experiment: a Dual-Dorsal model. We replaced the ventral stream with a second dorsal stream so that both pathways process low-resolution, grayscale dorsal-specific inputs. As shown in Table R1, neither pathway in the Dual-Dorsal model exhibits a monotonic increase in grid-like bases. This proves that simply scaling dorsal inputs to a dual-stream capacity does not induce an invariant topology; rather, the organic interaction between heterogeneous visual inputs (M- and P-pathways) designed in our study is required. Third, to directly prove that this property is not already built-in from initialization of training, we tracked the ratio of grid-like bases in the extended dorsal stream of ReGraph across training epochs (Table R2). As shown below, the monotonic pattern is absent at pre-training initialization and early training (Epoch 4), and progressively emerges and consolidates only as learning proceeds to optimize task performance. In conclusion, while our design choices provide the essential prerequisite foundation, the observed grid-like bases are not mere byproducts of dorsal-specific input; they genuinely emerge only through the integration of heterogeneous visual streams and task-driven learning.

Table R1. Layer-wise ratio of grid-like bases in the Dual-Dorsal model.

ModelStreamL1 (%)L2 (%)L3 (%)L4 (%)Average (%)Monotonic Trend
Dual-Dorsalstream-10.004.974.015.563.63X
stream-22.455.232.184.563.61X

Table R2. Evolution of grid-like bases in the dorsal stream of ReGraph across training epochs.

EpochTop-1grid-like bases ratio (%)monotonic trend
L1L2L3L4
At initialization—0.002.183.921.48X
epoch 462.6 %0.115.134.595.72X
epoch 970.1 %0.054.015.5011.75✓
epoch 1471.1 %0.112.944.7510.95✓
best (epoch 18)72.1 %0.004.818.9714.74✓
Reviewer_UPUgOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a dual stream architecture for the SSV2 benchmark. The architecture separates visual information into dorsal and ventral pathways, which are high-temporal-low-spatially and low-temporal-high-spatially resoluted, respectively. Each stream is processed through multi-head self attention, with optional dorsal to ventral flow.

The model achieved near SOTA performance on action classification. Grid patterns were found in the attention matrices. Gridness of the components are relevant to task performance. The gridness increases in the later two layers. Ablation of the ventral stream and dorsal-to-ventral modulation also showed drop in performance.

They conclude that the relational (grid) structure may arise not de novo, but rooted in the sensory divergence of the two pathways.

Critical strengths

The paper includes neuroscientific origins, interesting theoretical point of view, and good amount of experiments. The figures are well made and explains the methods with clarity.

Critical weaknesses

Poorly backed main claim

The paper did show how gridness can emerge by being trained on a task. But there is no evidence showing how the separation of dorsal and ventral pathways play necessary roles in it. Ablation of dorsal/ventral/D-V modulation removed the monotonic increase of gridness across layers, but not the gridness itself. The lack of gridness in other models does not offer a clean comparison for this claim, as the models differ in many aspects. The empirical results are interesting, but the framing is very misleading, and hides the true values of the results.

To go a bit further. I wonder what is the main message for neuroscience this paper conveys. Is monotonic increase of gridness across layers even biological? As far as I know, only entorhinal cortex shows reliable grid patterns across visual field. The monotonic increase is not a biological phenomenon to be matched to begin with.

SOTA models for SSV2 benchmark

The paper uses 33 classes out of the 174-class SSV2 benchmark. Although it's different from the full benchmark, it is still an unfair comparison with the known top performing models excluded (V-JEPA2 family, CAST, OMNIVORE, etc). I don't think the paper's model need to be the SOTA to be valuable, but it is misleading to give the impression of being SOTA.

Questionable analyses

Within the span of top 5% HOSVD components, the choice of bases are arbitrary (Lin et al., 2023). The singular vectors are simply ranked by singular value, not privileged for biological interpretations. Even for uniqueness' sake, a flatter singular value spetrum would result in a more arbitrary and noisier determination of modes. Simply choosing different linear combinations within the top 5% span would result in drastically different gridness scores. The shuffling do not eliminate this concern regarding the arbitrary choice of basis. One would need to make a method invariant to the choice of basis, or at least use some loading-sparsity based methods like in Lin et al. (2023) to identify interpretable components.

Unclear main contribution

As it stands, the paper does not have convincing proofs for either their neuroscience claim or their engineering advances, despite the interesting design and meticulous work. It needs a major reframing.

Questions for authors

Novelty

There is a long history of architectures mimicking the dorsal and ventral pathways. What is the novelty here? Is it the specific interpretation of dorsal and ventral pathways? Why not comparing this design to previous deisgns such as Thompson et al. (2024) Choi et al. (2023) and Daucé et al. (2020)?

15 comments

Dear Reviewer UPUg,

We sincerely thank the reviewer for the insightful comments and constructive feedback. Below, we have done our utmost to address your questions in detail. We hope our responses have addressed your questions and concerns, and would welcome any further discussion or requests for clarification.

(1) W2: "SOTA models for SSV2 benchmark." We would like to note that, as you correctly pointed out, our primary goal was not to establish a SOTA performance in the action recognition task but to examine how biologically motivated inductive biases, particularly the separation of dorsal and ventral processing streams, contribute to the emergence of grid-like relational representations. The performance comparisons with other algorithms were therefore intended to verify that ReGraph learns sufficiently discriminative representations to make the interpretation of the subsequent Relational Primitive Analysis (RPA) findings valid, not to claim superiority of our approach in the benchmark tests. Nevertheless, we agree that our comparisons should include more recent and general-purpose models to provide a broader contextualization of ReGraph as a new model in the AI field. Accordingly, we now additionally include V-JEPA2, CAST, and OMNIVORE in our comparison, as suggested by the reviewer. The table below summarizes their performance on SSv2, together with experimental settings such as pretraining datasets and parameter counts for fair interpretation. Importantly, we used publicly available checkpoints of CAST, V-JEPA2, and OMNIVORE, which were obtained through large-scale pretraining, and then subsequently fine-tuned them on the same 33-class training set used for ReGraph. Thus, the models were evaluated under largely identical conditions except for different pretraining conditions.

Modeltop-1SSV2 Pretrain# parameter
CAST84.68O217,120,161
V-JEPA288.56O353,255,457
OMNIVORE78.50X88,693,993
ReGraph74.57X90,236,420

(2) W3: "Questionable analyses." As the reviewer noted, gridness scores may vary depending on the basis selection in High-Order Singular Vector Decomposition (HOSVD), particularly when singular values are closely spaced. To assess the robustness of our findings, we performed two additional analyses using alternative basis-selection procedures. First, we tested whether the observed grid-like structure was driven by low-variance modes in the flat tail of the singular-value spectrum. For each condition, we identified the elbow (knee point) of the log-singular-value spectrum and retained only the prominent modes preceding this point, excluding the remaining tail modes. As shown in Table 1, the progressive increase in grid-like bases remained evident in ReGraph’s extended dorsal stream (0.00% → 24.06%), whereas the ventral and unmodulated streams did not exhibit the same monotonic pattern. Thus, the observed pattern persists even when less prominent spectral modes are excluded. Second, we examined whether the same pattern could be recovered using a different method for extracting components from the selected subspace. Following the reviewer’s suggestion, we applied the loading-sparsity framework of Lin et al. (2023), imposing an L1 penalty on the spatial dimensions of our attention matrices to identify sparse components. We selected the number of components (k) and sparsity penalty (α\alpha) through cross-validated grid search on held-out samples. This approach allowed us to examine whether grid-like organization was also present in components with more spatially concentrated loadings, beyond the HOSVD-derived bases. The resulting components also exhibited a progressive increase in grid-like structure across ReGraph’s extended dorsal layers (0.30% → 1.12% → 7.21% → 36.86%; Table 2), a pattern not observed in the other streams. The selected sparsity penalty ((α\alpha = 0.01)) provided mild regularization, reducing diffuse loadings while preserving spatially structured, multi-peak patterns. Together, the two analyses show that the progressive emergence of grid-like structure in ReGraph’s extended dorsal stream is not limited to our original, somewhat heuristic approach for HOSVD basis selection. The pattern remains evident both when low-variance modes are excluded and when components are extracted using a sparsity-constrained approach. Although these analyses do not 100% establish the uniqueness of individual basis vectors, they demonstrate at least that the main finding is robust to alternative basis-selection procedures.

Table1. Layer-wise Ratio of Grid-like Bases across Streams (%)

ModelStreamL1L2L3L4
ReGraphDorsal0.00%1.20%13.37%24.06%
Ventral0.00%10.33%6.01%10.58%
UnmodulatedDorsal0.00%7.45%4.23%3.97%
Ventral0.00%10.10%6.49%5.74%

Table2. Grid-like Base Ratio (%) under Loading-Sparsity (Lin et al., 2023 SCA)

ModelStreamL1L2L3L4
ReGraphDorsal0.30%1.12%7.21%36.86%
Ventral0.12%8.18%3.57%3.85%
UnmodulatedDorsal0.10%0.17%6.09%0.00%
Ventral0.10%15.58%6.86%1.44%

(3) Q1: "Novelty."

We agree that Thompson et al. (2024), Daucé et al. (2020), and Choi et al. (2023) propose compelling dorsal/ventral architectures for modeling cortical visual functions. However, fundamental differences between these and ReGraph in terms of (1) the research objective and (2) the specific architectural and input characteristics impose uncontrollable constraints on conducting a direct 1:1 comparison or uniformly applying our Relational Primitive Analysis (RPA) framework.

1. Divergence in Research Objectives and Preservation of Spatial Topology

While the three prior studies primarily model early-to-intermediate visual cortical functions (e.g., active vision, eye movements, or fMRI alignment) mainly through feedforward connections, ReGraph computationally tests a distinct systems-neuroscience hypothesis: Does the origin of grid-like patterns emerging in and upstream of the medial entorhinal cortex (MEC) lie in the information factorization of the dual visual streams? Although sharing a dual-stream paradigm, our distinct objective requires jointly implementing cortical feedforward, dynamic lateral, and recurrent connectivities alongside top-down dorsal-to-ventral (D→V) modulation. Crucially, to evaluate whether spatial bases exhibit hexagonal gridness across layers, ReGraph is explicitly designed to preserve the 2D spatial topology (N spatial nodes) throughout the hierarchy, unlike other models mentioned in the comment.

2. Structural Incompatibility with RPA (Thompson et al., 2024; Daucé et al., 2020)

Our RPA pipeline extracts dominant spatial bases via HOSVD from N × N node-to-node spatial interaction matrices. However, Thompson et al. (2024) and Daucé et al. (2020) structurally preclude this analysis: In Thompson et al. (2024), the dorsal stream does not spatially partition visual frames; instead, it receives only the 2D gaze coordinates (x,y) at the current fixation point and compresses them through fully connected layers into a single 1D latent vector (512 dimensions). Similarly, Daucé et al. (2020) feeds 2,880 static retinal filter response coefficients as a 1D vector into multi-layer dense networks. In both models, passing through feedforward dense layers immediately collapses the 2D spatial geometry into a single 1D vector, eliminating the spatial nodes representing screen locations and leaving no spatial dimension to extract the N × N interaction matrix required for RPA.

3. Empirical Verification on Choi et al. (2023)

Among the three models, only Choi et al. (2023) directly processes visual frames via dual CNN backbones (WhereCNN and WhatCNN) while preserving 2D spatial structure. Embracing the feedback of the reviewer, we fine-tuned their official pretrained checkpoint on our SSV2 subset under our exact protocol and performed the identical layer-wise RPA (Appendix K).

[Table R1. Layer-wise RPA on Choi et al. (2023)]

MetricStreamL1L2L3L4
Grid-like Code Ratio (%)WhereCNN (Dorsal)0.850.000.001.71
WhatCNN (Ventral)0.000.000.001.71
Context-Invariance (Acc. %)WhereCNN (Dorsal)11.414.926.427.8
WhatCNN (Ventral)12.616.529.549.0

As shown in Table R1, Choi et al. (2023) exhibits no monotonic layer-wise increase in grid-like codes in either WhereCNN or WhatCNN, contrasting sharply with ReGraph's dorsal stream (0.0% → 5.1% → 8.8% → 15.3%; Table 3). Moreover, in OOD linear probing on SSV2—which demands spatiotemporal generalization—Choi et al. (2023) showed low overall accuracy (peaking at 27.8% in Dorsal and 49.0% in Ventral vs. ReGraph's 97.4% and 95.0% at L4; Table 2) and lacked the inter-stream dichotomy observed in ReGraph. Thus, while prior dual-stream architectures excel at modeling specific visual cortical functions and representational alignment, these additional experiments reconfirm that simulating emergent relational structures and generalization requires the specific synergy of inductive biases integrated in ReGraph: M/P asymmetry, lateral/recurrent connectivities, and top-down modulation.

Comments on the paper

8 comments

(6) W7: “Positioning relative to prior biological dual-stream architectures and TEM family.”

We thank the reviewer for highlighting these highly relevant conceptual precedents. We agree that contextualizing ReGraph alongside SlowFast and the Tolman-Eichenbaum Machine (TEM/TEM-t) strengthens the manuscript's architectural positioning. We acknowledge that SlowFast is one of key precedents in utilizing M/P-inspired spatiotemporal asymmetry for dual-stream architectures. However, we would also like to note that while SlowFast pioneers M/P-inspired temporal asymmetry via differing frame sampling rates and filter sizes, ReGraph diverges in two key design principles. First, instead of using structurally similar CNN pathways, ReGraph explicitly enforces functional divergence using stream-specific pre-trained encodings (e.g., depth, optical flow, and saliency for the dorsal stream; object-oriented processing for the ventral). Second, rather than relying on feedforward convolutional fusion between corresponding layers, ReGraph implements the biological "magnocellular advantage" via top-down, dorsal-to-ventral cross-attention. Ultimately, ReGraph’s goal is not merely optimizing action recognition, but investigating how these biases drive the emergence of context-invariant relational primitives.

We also appreciate the pointer to TEM and TEM-t, which elegantly proposed a theoretical model of structure (MEC) and content (LEC) factorization for generalization and its relationship with Transformer, de facto the most potential AI architecture currently. However, whereas TEM/TEM-t focus their modeling within the MEC/LEC circuitry itself, ReGraph asks a more ontological question about the origin of this functionality from the upstream pathways. We hypothesize that grid-like bases do not arise de novo in the medial temporal lobe, but are already observed in the upstream by the functional divergence of the dorsal and ventral visual pathways. Methodologically, while TEM-t quantifies grid-like codes from recurrent positional encodings during path integration, ReGraph extracts them from the network's dynamic routing structures (attention bases) during visual generalization tasks. We have expanded Appendix A (Related Works) to integrate these close conceptual precedents and clarify ReGraph's distinct computational objectives and architectural implementations.

(7) W1: "Mechanism of Grid-Like Bases for Generalization through attention routing and transmitted values?" (1/2)

We agree that it is important to clarify how this periodic geometry functionally supports generalization through attention routing and transmitted values. To establish this mechanism, we conducted a series of complementary analyses comparing the dorsal and ventral streams in ReGraph, based on the core assumption that the ventral visual stream primarily processes sensory contents, whereas the dorsal stream processes their motion and spatial relations.

First, we examined how much the information routing in each pathway is driven by the sensory content. For this, we extracted the representations from nodes (=neurons) of each ReGraph’s layer, measured their across-node feature similarities, and correlated it with the corresponding attention weights. If nodes with a similar representation also receive stronger attention from one another, this would suggest that information loaded on that routing likely contains the sensory content. In contrast, a weaker correspondence would suggest that the loaded information less represents the content and may convey something else, for instance, spatial relations.

To characterize how much such spatial relations are indeed processed by the information flow within each pathway, we next analyzed the connectivity structure of the attention matrices. To this end, we first detected so-called ‘hub’ nodes that connect strongly with other nodes, serving as a central information processor through which signal is integrated and distributed across the network. We then examined how broadly these hubs were distributed across the 2D-sheet cortical network (=hub dispersion) and whether their spatial arrangement showed eventually grid-like organization.

  • Hub dispersion (Normalized Participation Ratio; NPR): We quantified the hub dispersion using Participation Ratio (PR) of the leading eigenvector of the graph Laplacian, which indicates whether information flow is dominated by a few spatial nodes or shared across many nodes. We then normalized the PR by the number of nodes in each stream (NPR) to compare across different spatial resolutions, with NPR close to 0 or 1 indicating node participation either to only restricted nodes (=localized connectivity) or to diffuse areas in a more uniform fashion, respectively. In ReGraph, the dorsal stream showed broad node participation (NPR = 0.118), whereas the ventral stream remained highly localized (NPR = 0.008) (Table R1).

  • Topographic arrangement of nodes (marginal gridness): Having established the PR of the connectivity structure, we next examined whether the attention patterns also formed periodic spatial arrangements. To assess this, we summarized connectivity along the target and source axes using marginal mean vectors (average connectivity associated with each spatial node) and reshaped them into 2D maps. We then quantified the proportion of marginal maps exhibiting grid-like patterns above the gridness threshold. Notably, in ReGraph, the dorsal stream organized these broadly dispersed hubs into periodic grid-like lattices (23.3% / 29.3% target/source marginal gridness), whereas the ventral stream remained spatially unorganized (3.1% / 4.1%) (Table R1). Because an operator's leading principal bases mathematically align with the dominant structure of its marginal mean profile (Kim & You, 2023), this periodic hub organization explains the L4 surge in grid-like bases in ReGraph (Table 3).

(8) W1: "Mechanism of Grid-Like Bases for Generalization through attention routing and transmitted values?" (2/2)

Lastly, to determine how such diffuse, periodic connectivity patterns translated into functional differences in the transmitted signals between the two pathways, we first measured whether information exchange between spatial nodes was guided by visual-feature similarity, and then used layer-wise relevance propagation (LRP) to identify the nodes that contributed most to the model’s output and visualize their corresponding regions in the input frames.

  • Content-independent vs. content-loaded routing (ρ): To determine whether each pathway routed information according to the similarity of visual content, we measured the Pearson correlation (ρ) between pairwise input-feature similarity and attention weights. A positive correlation indicates content-loaded routing, whereas a weak or negative correlation indicates content-independent routing. ReGraph showed a clear asymmetry: the dorsal stream routes information independently of local feature content (=-0.06), whereas the ventral stream relies heavily on feature similarity (=+0.43) (Table R1).

  • Global vs. local frame coverage (Layer-Wise Relevance Propagation; LRP): We next examined which regions of the input frames corresponded to the nodes that actually contributed to the model’s action classification output. Using LRP (Ali et al., 2022), we identified the most relevant nodes, mapped them to their corresponding regions in SSV2 input frames, and quantified their spatial coverage (Appendix M, Fig. 7 in the revised manuscript). These are also viewable directly in the browser (no download): https://anonymous.4open.science/r/regraph-351D/rebuttal/rebuttal/README.md. ReGraph showed a strong pathway distinction: the dorsal pathway covered broad regions of the frame (26.0%→33.3%→38.2%→42.2%26.0\% \rightarrow 33.3\% \rightarrow 38.2\% \rightarrow 42.2\% across L1–L4), whereas the ventral pathway focused tightly on localized object regions (4.4%→4.6%→4.9%→4.8%4.4\% \rightarrow 4.6\% \rightarrow 4.9\% \rightarrow 4.8\%), yielding a large layer-wise D–V coverage divergence (ΔD−V:21.7→28.7→33.3→37.4 %p\Delta_{D-V}: 21.7 \rightarrow 28.7 \rightarrow 33.3 \rightarrow 37.4\text{ \%p}). Taken together, these findings suggest that ReGraph supports broad, content-independent information processing in the dorsal stream, while maintaining localized feature-dependent processing of objects in the ventral stream.

In sum, these results show how periodic geometry supports OOD transfer through attention routing and transmitted values. Through attention routing, periodic geometry organizes the dorsal stream into broadly dispersed (NPR = 0.118), spatially periodic hub lattices (23.3% / 29.3% marginal gridness) rather than collapsing onto specific object locations. Through transmitted values, because this periodic routing operates independently of local visual similarity (=-0.06), the signals aggregated and transmitted across these hubs convey content-invariant spatial relations across the global scene (42.2% LRP coverage) rather than object-specific appearance features. This emergent division of labor closely mirrors computational neuroscience accounts proposing that the dorsal/entorhinal system employs grid-like basis functions to capture content-invariant relational structure (Behrens et al., 2018; Whittington et al., 2020), while the ventral stream processes localized object features.

Table R1. Analysis of L4 Attention Topology and Processed Information in ReGraph

ModelStreamHub Dispersion (NPR)Marginal Gridness (T / S)Content Corr. (ρ)LRP Frame Coverage (L1 L2 L3 L4)
ReGraph (Ours)Dorsal0.11823.3% / 29.3%-0.0626.0% → 33.3% → 38.2% → 42.2%
Ventral0.0083.1% / 4.1%+0.434.4% → 4.6% → 4.9% → 4.8%
Coverage (D - V)ReGraph---21.7 → 28.7 → 33.3 → 37.4 %p
  • Kim, D., & You, K. (2023). PCA, SVD, and Centering of Data. arXiv preprint arXiv:2307.15213.
  • Constantinescu, A. O., O’Reilly, J. X., & Behrens, T. E. (2016). Organizing conceptual knowledge in humans with a gridlike code. Science, 352(6292), 1464-1468.
  • Moser, E. I., & Moser, M. B. (2008). A metric for space. Hippocampus, 18(12), 1142-1156.
  • Ali, A., Schnake, T., Eberle, O., Montavon, G., Müller, K. R., & Wolf, L. (2022, June). XAI for transformers: Better explanations through conservative propagation. In International conference on machine learning (pp. 435-451). PMLR.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.