Reproduce “Linear-Core Surrogates: Smooth Loss Functions with Linear Rates for Classification and Structured Prediction” 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
Mehryar Mohri, Yutao Zhong
Linear-Core constructs convex classification losses with a straight, nonzero-slope region near the decision boundary and smooth tails. It proves linear H-consistency bounds for binary, multi-class, and structured prediction, subject to hypothesis-class conditions. For structured prediction, sampled gradients avoid enumerating label pairs. In this CIFAR-10 test, it reaches 75.49% accuracy under 40% instance-dependent label noise, versus 72.93% for tuned GCE.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.