Reproduce “Neural Classifiers as Error-Correcting Codes” 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
Sergey Salishev, Anton Makarov, Oleg Granichin, Renata Avros, Zeev Volkovich
The authors frame classification as error correction: classes are messages, and features are implicit codewords. Under stated assumptions, linear softmax decoding can act as a maximum-likelihood decoder. In a synthetic test, its accuracy tracks the optimal decoder with one codeword per class at moderate noise, but agreement weakens with noise and multiplicity; this does not establish equivalence for real networks or quantization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.