Reproduce “Depthwise Convolution is All You Need for Learning Multiple Visual Domains” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yunhui Guo, Yandong Li, Rogerio Feris, Liqiang Wang, Tajana Rosing

A multi-domain learning approach using depthwise separable convolutions implements a parameter sharing strategy where pointwise filters are shared globally and depthwise filters are domain-specific. The method achieved a score of 3507 on the Visual Decathlon Challenge, exceeding prior top methods by 95 points while requiring 50% fewer parameters.

Key insights

  • Cross-channel correlations are proposed to be universal across visual domains, whereas spatial correlations are largely domain-specific.
  • Network dissection demonstrated that depthwise convolutions capture higher-level semantic features, while pointwise convolutions primarily detect low-level texture-like attributes.
  • Sharing pointwise convolutional filters (cross-channel correlations) and specializing depthwise filters (spatial correlations) leads to superior multi-domain learning performance.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.