Reproduce “Zonotopic Analysis of Quantization Error in Neural Networks” 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
Robert Debbas, Éric Goubault, Sylvie Putot
The authors present a method to bound how quantization changes network outputs, abstracting rounding and clamping and propagating zonotopes through the original and quantized networks. On 10 ACAS Xu networks, at radius 3, constrained zonotopes achieve average error-interval width 1.4 and runtime 94 ms, versus 8.5 and 422 ms for symbolic DRA; at larger regions, symbolic DRA can be tighter.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.