Reproduce “Compacting Neural Network Classifiers via Dropout 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
Yotaro Kubo, George Tucker, Simon Wiesler
The authors introduce dropout compaction, which learns unit-specific retention probabilities and uses a prior favoring values near zero or one, allowing units to be removed during training. On VoiceSearchLarge, a model with roughly half the baseline’s hidden units matched its 18.9% word error rate (WER), the percentage of reference words requiring substitution, deletion, or insertion; annealed dropout scored 18.7%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.