Reproduce “Soft Threshold Weight Reparameterization for Learnable Sparsity” 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
Aditya Kusupati, Vivek Ramanujan, Raghav Somani, Mitchell Wortsman, Prateek Jain, Sham Kakade, Ali Farhadi
The authors introduce STR, a method that learns a separate cutoff for each neural-network layer. During the forward pass, STR zeros weights below the cutoff and shrinks larger weights, while backpropagation learns them. On ImageNet-1K, MobileNetV1 at about 89% sparsity reaches 62.10% top-1 accuracy with 42M FLOPs, versus GMP’s 61.80% and 82M; these are operation counts, not measured runtime.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.