Reproduce “An Extended Study of Human-like Behavior under Adversarial Training” 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
Paul Gavrikov, Janis Keuper, Margret Keuper
This study comprehensively investigates how adversarial training influences human-like perception and out-of-distribution generalization in both convolutional neural networks and Vision Transformers. It finds that adversarial training increases shape bias across architectures but often degrades performance on other OOD datasets, providing a frequency-domain explanation for these trade-offs.
Key insights
- Adversarial training consistently induces a more human-like shape bias in both CNNs and Transformers by desensitizing models to high-frequency information that is often exploited by attacks.
- The shift towards shape bias comes with a trade-off, as AT frequently leads to decreased generalization performance on OOD datasets whose frequency spectra significantly diverge from the clean training data.
- Transformers, even smaller variants, demonstrate superior OOD generalization and generally more human-like behavior compared to CNNs, a characteristic that largely persists under adversarial training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.