Reproduce “Attention over noisy examples: generalization and memorization in the exponential Hopfield network” 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
Francesco Alemanno
The paper tests whether attention can recover archetypes from noisy examples. It treats examples as unlabeled keys and values, using a softmax-weighted sum followed by a sign. At intermediate temperatures, attention pools same-class examples and suppresses interference. Structured-data test accuracy reaches 0.90 on MNIST and 0.80 on Fashion-MNIST, but the theory covers a single noiseless update and becomes less accurate after condensation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.