Reproduce “Truly shift-invariant convolutional neural networks” 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
Anadi Chaman, Ivan Dokmanić
Researchers from the University of Illinois at Urbana-Champaign and the University of Basel introduce Adaptive Polyphase Sampling (APS), a novel downsampling scheme that replaces conventional linear downsampling in convolutional neural networks to achieve true shift invariance. This architectural modification results in 100% classification consistency to arbitrary input shifts and generally improves classification accuracy, even on out-of-distribution data.
Key insights
- The root cause of CNNs' shift sensitivity lies in the fixed-grid nature of downsampling layers, which capture different pixel sets for shifted inputs, rather than in convolution or pooling operations.
- Non-linear activation functions significantly hinder the effectiveness of anti-aliasing methods by disproportionately thresholding subtly different feature maps, thereby reintroducing shift sensitivity.
- By adaptively selecting from all polyphase components, shift invariance can be intrinsically embedded into the network architecture, making it robust to shifts regardless of training data or subsequent non-linearities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.