Reproduce “MentorNet: Learning Data-Driven Curriculum for Very Deep Neural Networks on Corrupted Labels” 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

Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, Li Fei-Fei

MentorNet trains an image classifier on partly corrupted labels by learning weights that control how strongly each example influences a StudentNet, using label information and signals from the student’s progress. On CIFAR-10 with 40% randomly replaced labels, the data-driven version reaches 89% clean-validation accuracy versus 77% for a predefined curriculum; it also uses 5,000 clean-labeled examples, unlike that baseline.

Summary by alphaXiv.

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