Reproduce “PROFIT: A Novel Training Method for sub-4-bit MobileNet Models” 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
Eunhyeok Park, Sungjoo Yoo
The authors introduce PROFIT, which stabilizes low-bit MobileNet fine-tuning by measuring activation shifts across layers and freezing the most sensitive weights in stages. DuQ and negative padding address asymmetric h-swish activations. On ImageNet, 4-bit MobileNet-v3 reaches 73.812% top-1, versus 75.296% for Full+; MobileNet-v1/v2 remain within 0.5 top-1 points of baseline. Results reflect the training recipe, not measured device savings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.