Research map / Label-efficient and robust learning / Recognition and long-tail

Deep Edge Filter: Return of the Human-Crafted Layer in Deep Learning

NeurIPS 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Correspondence and causal representation learning · 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.

Proposes Deep Edge Filter, which subtracts a low-pass filtered version of intermediate deep features (high-pass residual) inserted between network blocks, and reports gains on test-time adaptation, GLUE, few-shot NeRF, and audio classification.

Problem

Deep networks generalize poorly under domain shift and perturbations, partly because they encode domain-specific low-frequency biases in deep features; classical edge filters on inputs lose fine detail and are limited to images.

Approach

Apply a high-pass residual (feature minus mean/median/Gaussian low-pass output) directly to intermediate deep features as a single non-learnable inserted layer, using 2D kernels for CNNs and 1D kernels over tokens/rays for transformers and MLPs, with the LPF branch detached.

Key results

• On CIFAR-10C with WRN28-10, Direct accuracy rises from 49.6% to 57.7% (+8.0) with the Edge Filter, and TENT rises from 74.6% to 75.8%.
• On SST-2 with BERT, accuracy improves from 79.36% to 80.85% (+1.49), and GLUE average improves by +0.81.
• On few-shot NeRF (8 views, Blender), average PSNR rises from 22.95 to 23.39 and LPIPS falls from 0.126 to 0.112, though ficus drops in PSNR by 0.21.

Measured effects reported in the paper

Surprising findings

• Replacing the Edge Filter with a trainable convolution of equal computation reduced corrupted-data accuracy (Direct 46.9 vs 49.6 baseline), suggesting the gain is not just extra compute.
• Applying the mean low-pass filter directly to features (instead of the high-pass residual) substantially degraded performance across tasks, e.g. -9.95 on WRN CIFAR-10C and -26.68 on ViT.

Limitations

• Experiments were limited to small or older models (WRN28-10, ResNet18, ViT-B/32, BERT); LLMs and state-of-the-art models were not tested.
• Models must be retrained from scratch with the filter attached, which limits applicability.
• Gains are inconsistent across configurations; e.g. ViT degrades with early-layer filters and Gaussian LPF, and the audio ablation fluctuates from -2.0 to +3.6 points.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 ImageNet200 CIFAR-10C CIFAR-100C ImageNet200-C GLUE SST-2 GLUE QQP GLUE QNLI Blender (NeRF synthetic) UrbanSound8K LAION2B

Benchmarks: CIFAR-10C CIFAR-100C ImageNet200-C GLUE SST-2 GLUE QQP GLUE QNLI Blender few-shot NeRF UrbanSound8K PACS DomainNet

Metrics: accuracy PSNR SSIM LPIPS MAE activation density

Baselines: NORM TENT Direct (no adaptation) NeRF LPF-only variant trainable Conv layer control

Builds on: Tent NORM Neural Radiance Fields BERT

Closest papers by meaning

• CNN Filter DB: An Empirical Investigation of Trained Convolutional Filters (CVPR 2022 · similarity 0.90)
• RePr: Improved Training of Convolutional Filters (CVPR 2019 · similarity 0.88)
• The Master Key Filters Hypothesis: Deep Filters Are General (AAAI 2025 · similarity 0.87)
• Deep Learning as Neural Low-Degree Filtering: A Spectral Theory of Hierarchical Feature Learning (NeurIPS 2026 · similarity 0.87)
• PICNN: A Pathway towards Interpretable Convolutional Neural Networks (AAAI 2024 · similarity 0.86)
• Data Filtering Networks (ICLR 2024 · similarity 0.85)