Reproduce “Learning Associative Inference Using Fast Weight Memory” 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

Imanol Schlag, Tsendsuren Munkhdalai, Jürgen Schmidhuber

Fast Weight Memory (FWM) gives an LSTM a rapidly editable memory that stores associations, revises them when addresses recur, and repeatedly reads retrieved values into later queries. This supports chained retrieval across a sequence. On constructed concatenated-bAbI question-answering, FWM reaches 96.75% ± 0.05 versus MNM’s 88.97% ± 6.28, but the benchmark is synthetic and memory is computationally demanding and size-limited.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.