Reproduce “Double-Scoring: Reliable Extraction of Strong Lottery Tickets” 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
Bryce A. Christopherson, Jack Baretz, Darian Colgrove, Salah Dandan
Double-scoring introduces a method to reliably extract "strong lottery tickets"—sparse subnetworks performant at initialization—from neural networks by dynamically learning layerwise sparsity. It achieves higher strong-ticket accuracy, reaching 82.74% on FashionMNIST at 90% requested sparsity, while eliminating the hyperparameter sensitivity of prior extraction techniques.
Key insights
- A fixed mask density of 1/2 in an appropriately augmented score space is sufficient to represent all possible subnetworks on original weights, addressing the sparsity selection bottleneck.
- Auxiliary scores act as a competitive mechanism, allowing the optimization process to implicitly determine effective sparsity levels on the original weights without explicit parameterization.
- The improved reliability in strong lottery ticket extraction primarily stems from a more effective search parameterization, bridging the gap between theoretical existence and practical recoverability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.