PaperlayerView exact revision on arXiv

Source-linked AI summary

The Shape of Events: Edge-Based Inductive Biases via Cross-Domain Distillation

Soshun Kihara, Shunsuke Yasuki, Masato Taki

arXiv:2609.30478v1cs.CV

TL;DR

CNNs’ texture bias and the lack of diagnostic benchmarks in the event domain leave event-induced model biases underexplored. The paper distills event knowledge into RGB models for analysis, finding color invariance, shape bias, and high-frequency-noise resistance associated with stronger reliance on edge-based shape.

  • Problem

    Event-domain inductive biases remain underexplored because diagnostic datasets and metrics for analyzing model behavior are lacking.

    Source: §1 · p. 2

  • Method

    The authors distill an N-ImageNet event teacher into an RGB ImageNet student so RGB-domain diagnostics can analyze event-derived representations.

    Source: §1 · p. 2, §3.1 · p. 4

  • Results

    Event-distilled models exhibit color invariance, shape bias, and high-frequency-noise resistance, with reduced texture dependence and stronger edge-based shape reliance.

    Source: §1 · p. 2, §8 · p. 10

  • Takeaways & Limitations

    The paper presents event-derived edge-based shape as an inductive bias that differs from existing robust models and supports downstream tasks.

    Source: §1 · p. 2, §8 · p. 10

    The evaluation is confined to CNNs and image classification, and N-ImageNet’s recaptured RGB images do not test high dynamic range or motion parallax.

    Source: §7 · p. 10

Abstract

from arXiv · show

Convolutional neural networks trained on ImageNet are known to exhibit a strong preference for local high-frequency texture, an inductive bias that translates into fragile robustness against distribution shifts in real-world environments. Event cameras, in contrast, record only changes in scene brightness and are therefore well suited to capturing contour information; however, due to the absence of diagnostic benchmarks in the event domain, the inductive bias that event-camera data instills in vision models has remained underexplored. In this work, we use knowledge distillation from the event domain to the RGB domain so as to exploit the rich evaluation toolkit available in the RGB domain and systematically dissect this inductive bias. Our experiments show that distillation from the event domain induces, in the RGB domain, color invariance, shape bias, and robustness to high-frequency noise. We identify the underlying mechanism as the model suppressing its dependence on high-frequency texture while acquiring a stronger dependence on edge-based object shape. This hypothesis is supported by changes in how color and spatial information are processed at the early layers, together with a spectral trade-off in which robustness to the absence of high-frequency components coexists with vulnerability to contamination of the relied-upon frequency bands and to disruption of geometric structure. We further show that this inductive bias differs from existing robustification methods and that it functions as a useful prior for diverse downstream tasks in which shape and contour information contribute alongside other cues. The code is available at https://github.com/snskysk/event2rgb-distillation .

1 Introduction

CNNs’ preference for local high-frequency texture is linked to fragility under environmental change, while event-domain representations remain underexplored because diagnostic benchmarks are lacking. The paper uses event-to-RGB distillation to study the resulting bias with RGB diagnostics, finding greater color invariance, shape bias, and resistance to high-frequency noise.

Motivation

  • CNNs’ preference for local high-frequency texture over global object shape is linked to fragility under illumination variation, noise, and domain shift.The paper frames this as a contrast with human vision [1].Source: §1 · p. 1

Research gap

  • Event-domain representations are underexplored because diagnostic metrics and datasets for analyzing model behavior are lacking in that domain.The RGB domain has established diagnostics, including ImageNet-C [3] and shape–texture cue-conflict images [1].Source: §1 · p. 2

Approach

  • The authors distill an N-ImageNet event teacher into an RGB ImageNet student to analyze event-derived features with RGB-domain evaluations.Because N-ImageNet recaptures ImageNet on a monitor, the experiments isolate brightness-gradient selectivity, not HDR or motion parallax.Source: §1 · p. 2

Main contributions

  • Event-distilled models show color invariance, a shift toward geometric and structural cues, and improved accuracy under high-frequency noise and small adversarial perturbations.The paper attributes these properties to reduced dependence on high-frequency texture and stronger dependence on edge-based shape.Source: §1 · p. 2

2 Related Work

The related work situates this study at the intersection of event-based vision, CNN shape and texture biases, and knowledge distillation. Its distinguishing approach sends event-model knowledge into RGB models to analyze event-induced representations.

2.1 Event-based Vision and Neuromorphic Datasets

  • Event-based vision work includes recognition models such as DiST [5] and growing applications in detection [33] [34] [35] [38] [39] and segmentation [40] [41] [42] [43] [44].Many methods convert event streams to grid representations so existing CNN architectures can be reused.Source: §2.1 · p. 2

2.3 Knowledge Distillation for Model Analysis

  • Unlike prior distillation from RGB into event models [45] [46] [47] [48], this work transfers event-model knowledge into RGB to analyze event-induced representations.The authors repurpose distillation for analysis rather than model compression or accuracy on a new task.Source: §2.1 · p. 2, §2.3 · p. 3

2.2 Inductive Biases in CNNs: Texture vs. Shape

  • Prior shape-bias approaches include Stylized-ImageNet [1], adversarial training, and noise augmentation; the paper distinguishes its edge-based shape feature from these methods.The claimed distinction is examined against Adversarial Training, Noise Training, and SIN [1].Source: §2.2 · p. 3

3 Method: Distilling Event-Driven Inductive Biases

The method repurposes cross-domain knowledge distillation to transfer event-domain predictions into an RGB student, making event-derived inductive biases accessible to RGB diagnostics. ResNet-34 is the principal architecture, with hard and soft labels mixed to preserve classification capability while transferring the teacher’s predictions.

3 Method: Distilling Event-Driven Inductive Biases

  • The framework transfers predictions from a DiST ResNet-34 trained on N-ImageNet to an ImageNet-trained RGB ResNet-34 student for analysis of event-domain inductive biases.The paired samples let the RGB student receive distillation guidance from the event teacher.Source: §3 · p. 3

3.1 Model Design and Training Strategy

  • ResNet-34 is chosen to match the DiST teacher, fit the compute budget, and support comparison with prior work; findings are also checked on other architectures and ResNet-50.The authors describe ResNet-34 as a classical CNN structure for this analysis.Source: §3.1 · p. 4
  • The student loss combines ground-truth cross-entropy with KL divergence to the teacher’s predictive distribution, weighted by α.The student uses paired RGB views and the corresponding event sample; the KL term uses temperature-softened logits with T = 3.0.Source: §3.1 · p. 4
  • ED_Student uses α = 0.2 to mix teacher soft labels with hard labels, while ED_SoftLabelOnly uses α = 1.0; the ratio is not exhaustively tuned.The α = 0.2 setting aims to preserve the event-derived bias while alleviating clean-accuracy degradation; the study prioritizes mechanism analysis over SOTA accuracy.Source: §3.1 · p. 4

4 Phenomena: Robustness Properties of Event-Distilled Models

RGB-domain evaluations find that event-distilled models maintain accuracy across hue changes, shift toward shape-based decisions, and resist high-frequency noise and small adversarial perturbations. These results are qualified by the limits of cue-conflict interpretation and the tested perturbation range.

4.1 Color and Hue Invariance

  • Across hue rotation angles, the Event-distilled model maintains color invariance, while grayscale evaluation shows smaller accuracy loss than the baseline.On grayscale ImageNet validation, the baseline changes from 73.9 to 62.4 and ED_Student from 69.7 to 64.6.Source: §4.1 · p. 5, Figure 2 · p. 5

4.2 High Shape Bias

  • Event-distilled models shift cue-conflict decisions toward shape and approach the human reference, unlike the baseline’s texture-based decisions.Patch-shuffle accuracy falls more sharply than the baseline’s as finer mosaics increasingly destroy image shape.Source: §4.2 · p. 5, Figure 2 · p. 5
  • The cue-conflict result indicates a shift in preference toward shape, not proof of strict texture non-reliance; patch-shuffle and patch-rotation probes are complementary.The authors cite Burgert et al. [73] on the interpretation of cue-conflict results.Source: §4.2 · p. 5

4.3 Robustness to High-Frequency Perturbations

  • Event-distilled models outperform the baseline under PGD perturbations with 0 < ϵ ≤ 1/255 and show greater resistance to high-frequency-band random noise.The paper does not evaluate larger adversarial budgets or stronger ensembles such as AutoAttack [76].Source: §4.3 · p. 5, Figure 2 · p. 5
  • The Figure 3 overview contrasts greater retention under JPEG compression with a high-severity reversal under pixelate corruption.The corruptions are reported as accuracy retention relative to clean performance.Source: Figure 3 · p. 6

4 Phenomena: Robustness Properties of Event-Distilled Models

  • The three robustness phenomena are presented as evidence of a shift in representation, which the paper investigates through internal representations and responses to artifacts.The authors use this analysis to seek a unified mechanism for the observed properties.Source: §4 · p. 4, §4.3 · p. 6

5 Mechanism: An Edge-Based Shape Feature Hypothesis

The paper hypothesizes that event-domain distillation shifts the model from unstructured high-frequency texture toward edge-based geometric structure. Filter analyses and robustness tests support this account while revealing sensitivity when edge continuity is disrupted.

5 Mechanism: An Edge-Based Shape Feature Hypothesis

  • Event distillation is hypothesized to suppress reliance on unstructured high-frequency texture while strengthening reliance on macroscopic geometric structure carried by edges.The paper frames this as the mechanism linking color invariance, shape bias, and resistance to high-frequency noise.Source: §5 · p. 6

5.1 Internal Representation: Structural Prior in First-Layer Filters

  • First-layer filters shift from noisy local-texture patterns toward geometric structures, including color-opponent-like boundaries and directional edge-extraction patterns.The authors argue these filters selectively respond to structural gradients rather than simply low-pass smoothing the input.Source: §5.1 · p. 6
  • ED_SoftLabelOnly filters show higher RGB and spatial correlations than baseline, with synchronous geometric extraction emphasizing continuity between adjacent pixels.The paper notes that aggregate correlations need not match the individual-filter patterns shown in Figure 3a.Source: §5.1 · pp. 6–7

5.2 Validation via ImageNet-C: Spectral and Structural Trade-offs

  • Under JPEG compression, Event-distilled accuracy retention exceeds the baseline even at high severities when macroscopic brightness gradients remain.This supports tolerance to removal of unstructured texture rather than general robustness to all frequency changes.Source: §5.2 · p. 7
  • At high pixelation severities, where edges become discontinuous, Event-distilled retention falls below baseline, despite resistance at low severities.The paper also reports similar sensitivity under contrast and elastic_transform corruptions.Source: §5.2 · p. 7

5.3 Validation via Information Removal: Dominance of Geometric Features

  • Across B&W, Canny-edge, and silhouette inputs, the baseline degrades sharply while ED_Student remains above it, consistent with an inference shift from texture to geometry.ED_Student also degrades under these shape-only conditions.Source: §5.3 · p. 7

6 Validation and Practical Implications

The validation tests whether event distillation’s effects come from color removal, generalize across architectures, transfer to downstream tasks, and differ from non-event supervision. ED_Student is the only evaluated model improving over baseline on all four reported axes, though its adversarial advantage does not persist against a stronger baseline.

6.2 Architectural Generality

  • ED_Student improves over baseline in 19 of 30 tested architecture, attack, and perturbation settings at ϵ ≤ 1/255, with most exceptions in ConvNeXt-Base and RepLKNet31B.Across architectures, shape-bias tests also reproduce the shift toward shape and a steeper accuracy drop as the grid becomes finer.Source: Figure 5 · p. 8, §6.2 · p. 8

6.1 Ablation Study: Is it merely the absence of color?

  • Against Self, Gray, and B&W distilled controls, Event-distilled models resist adversarial attacks better, indicating the transferred bias is not merely color removal.The paper attributes the difference from B&W distillation to sensitivity to brightness gradients characteristic of event sensors.Source: §6.1 · p. 8

6.3 Transferability: Event Distillation as a Structural Prior

  • Despite 4.2 percentage points lower pre-training ImageNet accuracy, ED_Student outperforms baseline on many downstream tasks under linear-probe transfer.The largest gains are reported on CIFAR-100/CIFAR-10, Chest X-ray, and MNIST, with gains also on ImageNet-Sketch and cars196.Source: §6.3 · p. 9, Table 1 · p. 9

6.4 Necessity: What Non-Event Supervisions Fail to Reproduce

  • Canny- and Synth-distilled controls reproduce different subsets of the profile, while adversarial training also gains on adversarial and high-frequency measures without reproducing the full joint profile.The authors therefore do not claim any single property requires event data; the examined non-event supervisions fail to reproduce the combination.Source: §6.4 · pp. 9–10
  • ED_Student is the only one of six models improving over baseline on all four axes: color, shape, high-frequency robustness, and downstream transfer.The axes are defined as hue-rotation accuracy, ImageNet-Sketch accuracy, reduction in band-limited-noise accuracy drop, and mean linear-probe accuracy.Source: Table 2 · p. 9, §6.4 · p. 10
  • With a stronger timm-style baseline, ED retains advantages on shape (+2.0 pp), color (+1.8 pp), high-frequency-band robustness (−4.2 pp drop), and transfer (+1.3 pp), but not adversarial robustness.The stronger augmentation already makes the SOTA baseline robust at the tested perturbation sizes.Source: §6.4 · p. 10

7 Limitations and Future Work

The authors bound their claims to brightness-gradient selectivity in a specific recaptured-image setting. Their evaluation is also limited to CNNs and image classification, and the reported effect diminishes with model capacity.

7 Limitations and Future Work

  • Because N-ImageNet recaptures RGB images from a monitor, its setup does not exercise high dynamic range or motion parallax; claims are restricted to brightness-gradient selectivity.The source images are planar, so the setup does not include motion parallax.Source: §7 · p. 10
  • The effect diminishes with model capacity: ConvNeXt-Base and RepLKNet31B gain less than ResNet-34 and account for most architectural exceptions.Source: §7 · p. 10
  • Evaluation is confined to CNNs and image classification; proposed extensions include Vision Transformers, dense prediction, feature-level distillation comparisons, and fine-tuning with Eq. 1.Source: §7 · p. 10

8 Conclusion

Cross-domain distillation is used to analyze what event-domain supervision transfers into RGB representations: the student suppresses unstructured high-frequency texture dependence and relies more strongly on edge-based shape. The appendix documents training and evaluation choices, including a clean-accuracy trade-off and a cross-modal view mismatch that was not experimentally isolated.

8 Conclusion

  • Event-distilled models suppress unstructured high-frequency texture dependence and rely more strongly on edge-based shape, with smoother early filters and a spectral trade-off.Robustness to texture loss coexists with vulnerability to frequency-band contamination and geometric disruption.Source: §8 · p. 10

A.1 Datasets

  • N-ImageNet event samples are paired instance-by-instance with corresponding ImageNet RGB images, enabling cross-domain distillation on matching examples.Both datasets have the same 1,000-class structure.Source: §A.1 · p. 16

A.3 Student Training

  • Student training combines cross-entropy on strongly augmented RGB inputs with a KL term from the paired event teacher; the KL temperature is 3.0.Event inputs receive random time and horizontal flips and spatial shifts, while the student KD branch uses a deterministic weak RGB view.Source: §A.1 · p. 16, §A.3.1 · p. 18, §A.4.4 · p. 19

A.2 Teacher Models

  • The study uses a frozen publicly released DiST ResNet-34 event teacher, while additional DiST teachers are trained in-house for four other architectures.The public ResNet-34 checkpoint reaches 48.43% top-1 on N-ImageNet validation.Source: §A.2.1 · p. 16, §A.2.3 · p. 17

A.4.5 Cross-Modal Augmentation Alignment in the KD Branch

  • The teacher–student view mismatch was not isolated experimentally, although the authors argue the class-probability KL makes it a stochastic regularization signal rather than systematic label noise.The teacher receives a stochastically augmented event view while the student uses a deterministic weak RGB crop.Source: §A.4.5 · pp. 19–20

A.5 Downstream Linear-Probe Protocol and Hyperparameters

  • Downstream transfer is measured with a strict linear probe: the ResNet-34 backbone is frozen and only a new linear layer is trained.The protocol is intended to isolate pretrained-feature transfer rather than backbone fine-tuning capacity; absolute small-image baselines are protocol-bound.Source: §A.5 · pp. 20–21

A.6 Compute and Runtime

  • Reproducing all nine variants takes about 11 days across 11 runs on the reported hardware, while checking the central baseline-versus-ED_Student claim takes about two days.The authors release trained checkpoints and an ImageNet-100 configuration for lighter-weight replication.Source: §A.6 · p. 21

A.7 Summary of Trained Models and Clean Validation Accuracy

  • ED_Student improves on ED_SoftLabelOnly for every architecture but remains 3.10–13.11 percentage points below the corresponding baseline in clean accuracy.ED_SoftLabelOnly ranges from 24.14% to 70.97%, depending on the teacher–student pairing.Source: §A.7 · p. 22

A.7.1 Teacher Confidence and the Ablation Confounder

Across the compared methods, ED_Student is the only supervision reported to improve over baseline simultaneously on color, shape, high-frequency-band robustness, and downstream transfer. The appendix also tests teacher-confidence explanations, additional controls, a stronger training recipe, and hue-rotation behavior.

A.7.1 Teacher Confidence and the Ablation Confounder

  • The authors argue teacher entropy alone does not explain ED_Student’s directional signatures; ED_SoftLabelOnly instead shows the strongest reported signatures despite using only KL supervision.It has the highest filter correlation and most pronounced shape bias, contrary to the proposed entropy-confound prediction.Source: §A.7.1 · p. 22

A.8 Additional Non-Event Controls

  • Canny_Distilled and Synth_Distilled provide non-event controls, with the latter built from paired, transformed still-RGB views whose difference yields contour-following ON/OFF responses.The authors specify that Synth_Distilled is neither an event simulator output nor a real-camera output.Source: §A.8 · p. 23

A.9 SOTA Training Recipe

  • Under the stronger SOTA recipe, ED’s advantage persists on shape, color, high-frequency-band robustness, and downstream transfer, but not on adversarial axes.The table caption attributes the latter boundary to the stronger augmentation making the SOTA baseline robust at those perturbation sizes.Source: §A.9 · p. 23, Table 12 · p. 24

A.10 Cross-Method Comparison on the Figure-1 Axes

  • ED_Student is the only method reported to improve over baseline simultaneously on color, shape, high-frequency-band robustness, and downstream transfer.The comparison spans linear-probe transfer, hue, grayscale, cue-conflict, sketch, bandpass, and adversarial axes.Source: Table 13 · p. 24, §A.10 · p. 24

B.1 Across-Model Comparison under Hue Rotation

  • Adversarial Training has the largest hue-rotation accuracy drop, while SIN retains baseline-like hue robustness despite strong shape bias.The authors attribute the adversarial result partly to hue rotation being outside the small per-pixel L∞ perturbation model.Source: §B.1 · p. 25
  • Gray Distilled and B&W Distilled improve hue invariance over baseline, but their gains are comparable to or smaller than ED_Student’s.The authors argue that color removal alone cannot fully explain ED_Student’s invariance, which they link to dependence on event-domain brightness gradients.Source: §B.1 · p. 25

B.2 Robustness under Grayscale Inputs

Event-distilled models retain more accuracy when color is removed and show hue-invariant responses in early layers. Across shape diagnostics, they rely on macroscopic shape and edge continuity, with a more favorable evaluated trade-off than several comparison models.

B.2 Robustness under Grayscale Inputs

  • Under grayscale, ED_Student outperforms baseline, while Gray Distilled attains the highest accuracy; Adversarial Training falls roughly 13 pp below baseline.SIN remains at baseline level, so shape bias acquired by another route does not itself confer color invariance.Source: §B.2 · pp. 25–26, Table 14 · p. 26

B.3 Layer-wise Feature Maps under Hue Rotation

  • ED_SoftLabelOnly’s early-layer responses remain almost invariant to hue rotation, unlike baseline channels that vary with input hue.The reported response-level evidence places color invariance near the input, rather than in output post-processing.Source: §B.3 · p. 26, Figure 9 · p. 27, Figure 10 · p. 28

C Extended Analysis of Shape Bias

  • The appendix extends shape-bias analysis across architectures and models, shape-only conditions, patch disruptions, and first-layer filters.The supplied results show consistent shape-oriented behavior across these diagnostics, with stronger dependence on macroscopic shape continuity than several comparison routes.Source: C Extended Analysis of Shape Bias · pp. 26, 29, §C.1 · p. 29, §C.2 · p. 30

C.1 Cue-Conflict: Across Architectures and Models

  • On VGG-16 and ConvNeXt-Base, Event-distilled models shift toward shape decisions; on ResNet-34, shape-bias scores range from 0.375 for ED_Student to 0.487 for ED_SoftLabelOnly.Adversarial Training scores 0.572 and SIN 0.428; the authors report that Event-distilled models preserve color invariance and transfer performance while retaining strong shape bias.Source: §C.1 · p. 29, Figure 11 · p. 29

C.2 Robustness under Shape-Only Inputs

  • On shape-only inputs, SIN leads on sketch, B&W, and Canny, while ED_Student leads on silhouette; Event-distilled models reach comparable shape-based inference without explicit shape training.Adversarial Training falls below baseline on all four conditions, suggesting texture suppression alone does not ensure shape inference.Source: §C.2 · pp. 29–30

C.2.1 Generation of Canny and Silhouette Evaluation Inputs

  • Canny inputs are grayscale, blurred, and thresholded into binary edge maps, while silhouettes are binary foreground masks; both are saved as PNG for evaluation.These inputs use the same evaluation transform as RGB and grayscale columns.Source: §C.2.1 · p. 30

C.3 Patch-Shuffle and Patch-Rotation Experiments

Patch disruptions test whether models use local texture or shape spanning across patches. Event-distilled models lose accuracy more sharply as macroscopic continuity is disrupted, while their filters and frequency-noise responses provide complementary evidence about the underlying representation.

C.3 Patch-Shuffle and Patch-Rotation Experiments

  • As patches become finer, Event-distilled models lose accuracy more sharply than baseline under shuffling and rotation, indicating reliance on macroscopic shape continuity rather than within-patch texture.The same pattern appears across VGG-16, ConvNeXt-Base, ResNet-50, and RepLKNet31B; SIN’s curves are comparable to baseline.Source: §C.3 · pp. 30, 32, Figure 12 · p. 31, Figure 13 · p. 31

C.4 Functional Taxonomy of First-Layer Filters

  • Across all 64 first-layer filters, Event-distilled models—especially ED_SoftLabelOnly—show smooth, spatially correlated structures, including global monochromatic, color-opponent structural, and Gabor-like edge filters.ED_SoftLabelOnly has RGB Correlation 0.747 vs. 0.549 for baseline and Spatial Autocorrelation 0.575 vs. 0.391.Source: §C.4 · pp. 32–33, Figure 14 · p. 32, Table 15 · p. 33

C.4.1 Robustness of the Mechanistic Claim to the Augmentation Pipeline

  • The authors argue that removing RandAugment and RandomErasing alone is unlikely to explain ED_SoftLabelOnly’s smooth filters; they attribute filter organization to the loss target and N-ImageNet input statistics.ED_Student’s first-layer correlations remain near baseline, yet its behavioral signatures persist, so the proposed representation shift is not localized to first-layer smoothness.Source: §C.4.1 · p. 33

C.4.2 Where the Filter Change Is Localized

  • Event-supervised students are most distinct from other students at layer1, where ED_Student–ED_SoftLabelOnly similarity is 0.161; the difference is indistinguishable at layers 3 and 4.The authors distinguish this layer-wise similarity measure from the filter-correlation statistics aggregated within a layer.Source: §C.4.2 · pp. 33–34

D Extended Analysis of High-Frequency Robustness

  • The appendix extends high-frequency robustness analysis through PGD settings, band-limited noise, ImageNet-C corruptions, and frequency-based interpretation of additive-noise vulnerability.The supplied results show that robustness depends on whether perturbations affect high-frequency texture or the bands used for edge extraction.Source: D Extended Analysis of High-Frequency… · p. 34, §D.1 · p. 34, §D.2 · p. 35

D.1 PGD Evaluation Hyperparameters

  • The PGD comparisons use a shared evaluation pipeline with 10 steps, one random restart, and budgets up to 1/255; larger budgets are not reported because models collapse in this setting.The attack operates in [0, 1] pixel space, with normalization inside a wrapper.Source: §D.1 · p. 34

D.2 Frequency-Band Random Noise

  • At 0.95π noise, ED_Student accuracy drops 12.2 pp versus 29.6 pp for baseline, but both are affected; their gap narrows under low- and mid-frequency noise.The reported band-pass drops depend on the fixed protocol, including noise magnitude and mask settings.Source: §D.2 · pp. 34–35, Figure 15 · p. 35

D.3 Comprehensive Evaluation on ImageNet-C

ImageNet-C results qualify the Event-distilled model’s robustness: gains are clearest when high-frequency texture is removed while contours remain, whereas disrupting edges or their supporting frequency bands can erase that advantage.

D.3 Comprehensive Evaluation on ImageNet-C

  • Event-distilled models exceed baseline on JPEG compression and fog, but blur’s advantage is limited and can reverse at high severity as edges are attenuated.Under noise corruptions they perform about as well as baseline or worse; severe pixelation, contrast, or elastic transformation can also reverse the advantage when edge continuity is lost.Source: §D.3 · p. 35, Figure 16 · p. 36

D.4 Why Event-distilled Models are Fragile to Additive Noise

  • Additive noise contaminates Low–Mid frequency bands that carry edge information, explaining why Event-distilled models can be vulnerable despite resisting high-frequency noise.Canny edge maps also concentrate energy in Low–Mid bands, and the authors characterize the vulnerability as dependence on edge-carrying frequencies rather than fragility to noise in general.Source: §D.4 · pp. 36–37, Figure 17 · p. 36, Figure 18 · p. 37

E Spectral Statistics of Event vs. Natural Images

The appendix compares ImageNet and N-ImageNet frequency statistics. N-ImageNet has a flatter power spectrum, which the authors relate to event sensing and to the edge-oriented representations of distilled students.

E Spectral Statistics of Event vs. Natural Images

  • ImageNet has PSD exponent α = 2.46, while N-ImageNet has α = 1.48, indicating weaker low-frequency dominance and a relatively larger mid- to high-frequency proportion in N-ImageNet.Fits estimate the exponent over the intermediate band 0.02F ≤ f ≤ 0.25F; the authors connect the flatter spectrum to event sensors’ recording of local brightness changes, while noting that edge information also occupies low frequencies.Source: E Spectral Statistics of Event vs.… · pp. 37–38, Figure 19 · p. 38

F Extended Discussion of Limitations

The limitations concern the teacher data-generation process, the scope of model and task validation, and the need to test the findings with more native event datasets.

F Extended Discussion of Limitations

  • N-ImageNet events were generated by recapturing RGB images displayed on a monitor, which limits exploitation of the sensor’s intrinsic high dynamic range.The captured content is planar, so complex motion parallax from real-world depth is absent.Source: F Extended Discussion of Limitations · pp. 38–39
  • How Event distillation changes Vision Transformer representations remains unknown, since this validation targets CNNs’ texture bias and ViTs are reportedly more shape-biased.Source: F Extended Discussion of Limitations · p. 39
  • Validation is limited to image classification; testing more native event datasets and whether the edge representation helps dense prediction remain future research directions.Semantic segmentation is given as an example of a dense-prediction task where the representation may be a useful prior.Source: F Extended Discussion of Limitations · p. 39
Loading 2609.30478v1…