Sparsity level becomes a learned parameter in vision model training

single source· 1 articles · confidence: medium · first seen 2026-09-16 20:00 UTC

What this means for you

Nothing to act on. This is an arXiv preprint: no scores, no evaluation dates, no results beyond the abstract's claim that learning the sparsity dial beats fixing it by hand. If you work on robustness for vision models, it is a method to read, not a component to adopt.

A preprint posted on 16 September 2026 makes the sparsity setting in convolutional sparse coding — the dial controlling how much of an image is discarded as redundant — learned during training rather than fixed by hand. The authors unfold the optimisation with FISTA, an iterative solver unrolled into network layers, and frame the trade-off as an information bottleneck. A label-free post-training step raises compression for corrupted inputs while the main weights stay frozen. They report comparable clean accuracy on CIFAR and ImageNet and better robustness under perturbation, but no scores or evaluation dates.

Key facts

  • ·The paper is arXiv preprint 2609.19122, posted 16 September 2026. source
  • ·The sparsity coefficient is treated as a differentiable variable learned jointly with the network parameters, replacing a value that was previously fixed and manually selected. source
  • ·The convolutional sparse coding optimisation is unfolded using FISTA, so the iterative solver becomes part of the network. source
  • ·A label-free post-training strategy adjusts compression strength for corrupted inputs while the main network parameters stay fixed. source
  • ·Experiments were run on CIFAR and ImageNet; the abstract reports competitive clean-data recognition and improved robustness under input perturbations without giving scores or evaluation dates. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire