Reproduce “Measuring the tendency of CNNs to Learn Surface Statistical Regularities” 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
Jason Jo, Yoshua Bengio
Researchers at MILA and Université de Montréal found that Convolutional Neural Networks (CNNs) exhibit substantial drops in performance, up to 28% on CIFAR-10, when image Fourier statistics in test data differ from training, even when objects remain human-recognizable. This indicates CNNs heavily rely on surface statistical regularities rather than higher-level semantic understanding for generalization.
Key insights
- Deep CNNs exhibit a strong tendency to generalize by adapting to superficial statistical regularities (like Fourier image statistics) present in their training data, rather than learning truly abstract conceptual features.
- This reliance on surface statistics offers a quantitative explanation for the paradox of CNNs' high performance coupled with their extreme sensitivity to slight data perturbations, such as adversarial examples.
- Simply increasing the depth or capacity of CNN architectures does not fundamentally mitigate this tendency to latch onto statistical shortcuts, highlighting a limitation in current model designs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.