Reproduce “Shake-Shake regularization” 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
Xavier Gastaldi
Shake-Shake regularization enhances deep neural network generalization by stochastically blending intermediate feature maps in multi-branch architectures during both forward and backward passes. This method achieved new state-of-the-art single-shot test errors of 2.86% on CIFAR-10 and 15.85% on CIFAR-100, outperforming prior approaches.
Key insights
- The stochastic perturbation during the backward pass, involving the application of new random coefficients, is critical for the regularization's effectiveness.
- Applying regularization coefficients independently to each image within a mini-batch ("Image" level) consistently yields stronger regularization benefits compared to batch-level application.
- Shake-Shake regularization reduces the correlation between the output feature maps of parallel residual branches, indicating that it encourages the network to learn more distinct and diverse feature representations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.