Training-adaptive convolutional sparse coding boosts robustness via information bottleneck

SUAT-SZ · hf · 2026-09-21

A new paper proposes training-adaptive convolutional sparse coding (CSC) for robust visual representation. The authors unfold CSC optimization with FISTA and treat the sparsity coefficient — usually fixed and hand-tuned — as a differentiable variable learned jointly with network parameters, interpreted through the information bottleneck: the sparsity term promotes compact representations while reconstruction and task loss preserve task-relevant signal. A label-free post-training strategy further adjusts compression for corrupted inputs. Experiments on CIFAR and ImageNet show competitive clean accuracy and greatly improved robustness under perturbations.

Original post →

More from Research

Research channel →