Research map / Label-efficient and robust learning / Domain adaptation and generalization

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

NeurIPS 2026 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Augmentation and mixture-of-experts generalization · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Distills a frozen event-domain DiST ResNet-34 teacher (trained on N-ImageNet) into an RGB ResNet-34 student to use RGB diagnostics for dissecting the inductive bias event data induces, finding color invariance, shape bias, and high-frequency noise robustness.

Problem

Event-camera-trained vision models lack diagnostic benchmarks, so the inductive bias that event data instills in visual representations is underexplored; the paper asks what event-domain supervision transfers into RGB representations.

Approach

Use cross-domain knowledge distillation from an event-domain teacher to an RGB student (paired at instance level via N-ImageNet/ImageNet) so the rich RGB evaluation toolkit can dissect the event-derived inductive bias; the mechanism is an edge-based shape feature that suppresses high-frequency texture dependence.

Key results

• ED_Student (alpha=0.2) reaches 69.7% clean ImageNet top-1 vs 73.9% baseline, but improves hue-rotation worst-case accuracy (63.4 vs 56.7) and grayscale accuracy (64.6 vs 62.5).
• ED_Student improves over baseline in 19 of 30 (architecture, attack, epsilon) settings for epsilon <= 1/255 under FGSM and PGD across five architectures.
• In linear-probe transfer over 19 datasets, ED_Student has mean gain +1.6 pp over baseline and is the only method among six that improves on all four axes (color, shape, high-frequency, transfer).

Surprising findings

• ED_Student is the only evaluated method improving on all four axes at once, while Adversarial and Noise Training degrade downstream transfer.
• Clean ImageNet accuracy drops 4.2 pp for ED_Student despite gains on shape, color, and transfer tasks.

Limitations

• N-ImageNet is produced by monitor recapture, so HDR and motion parallax are not exercised; claims limited to brightness-gradient selectivity.
• Effect diminishes with model capacity (ConvNeXt-Base and RepLKNet31B gain less and account for most exceptions).
• Evaluation confined to CNNs and image classification; logit-only KD without feature-level comparison.

Experimental setup

Datasets: ImageNet N-ImageNet Stylized-ImageNet ImageNet-C cue-conflict images ImageNet-Sketch CIFAR-100 CIFAR-10 MNIST Chest X-ray

Benchmarks: ImageNet validation ImageNet-C cue-conflict ImageNet-Sketch 19-task linear probe transfer suite PGD/FGSM adversarial attacks

Metrics: top-1 accuracy shape bias proportion accuracy retention delta score over baseline

Baselines: baseline ResNet-34 Adversarial Training Noise Training SIN+IN co-training Canny_Distilled Synth_Distilled Self Distilled Gray Distilled B&W Distilled

Builds on: knowledge distillation DiST Stylized-ImageNet DeiT Patient Teacher

Closest papers by meaning

• ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression (NeurIPS 2025 · similarity 0.86)
• Visual Anagrams Reveal Hidden Differences in Holistic Shape Processing Across Vision Models (NeurIPS 2025 · similarity 0.86)
• E2HQV: High-Quality Video Generation from Event Camera via Theory-Inspired Model-Aided Deep Learning (AAAI 2024 · similarity 0.85)
• Re-coding for Uncertainties: Edge-awareness Semantic Concordance for Resilient Event-RGB Segmentation (NeurIPS 2025 · similarity 0.85)
• EventPillars: Pillar-based Efficient Representations for Event Data (AAAI 2025 · similarity 0.85)
• AE-NeRF: Augmenting Event-Based Neural Radiance Fields for Non-ideal Conditions and Larger Scenes (AAAI 2025 · similarity 0.85)