Reproduce “Understanding Batch Normalization” 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
Johan Bjorck, Carla Gomes, Bart Selman, Kilian Q. Weinberger
Research from Cornell University empirically demonstrates that Batch Normalization's primary benefit is enabling deep networks to train with larger learning rates by preventing activation and gradient explosion, rather than solely alleviating internal covariate shift.
Key insights
- Batch Normalization's primary benefits, including faster convergence and improved accuracy, stem from its ability to enable larger learning rates.
- BN mitigates the explosion of activation magnitudes and uncontrolled gradient growth, which typically cause unnormalized deep networks to diverge at high learning rates.
- Standard random initialization schemes lead to ill-conditioned weight matrices in deep unnormalized networks, resulting in input-independent and poorly behaved gradients.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.