Reproduce “Learning Computational Structure for Sample-Efficient Prediction” 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
Hangyu Zhou, Constantine Dovrolis, Stephen Mussmann
LeWi learns neuron positions in a low-dimensional space and uses distances to determine which neurons in adjacent layers connect, jointly training positions and weights. Across 32 sampled depth-4 computation graphs, mean learning-curve area was 0.197 versus 0.268 for a dense network, a reported 27% reduction. This indicates better sample efficiency on tested compositions, but not on a high-arity control.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.