Reproduce “PARQ: Piecewise-Affine Regularized Quantization” 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

Lisa Jin, Jianhao Ma, Zechun Liu, Andrey Gromov, Aaron Defazio, Lin Xiao

PARQ trains low-bit vision models with a convex piecewise-affine regularizer and gradually hardens quantization during training. On ImageNet classification, 1-bit DeiT-S achieves 71.06 ± 0.02% accuracy, versus 70.07 ± 0.03% for STE/BinaryConnect and 70.69 ± 0.07% for BinaryRelax. Only attention-block weights are quantized; the result is not universal, and deployment efficiency is unmeasured.

Summary by alphaXiv.

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