Reproduce “Self-Supervised Collapse Prevention with Score-Based Entropy Maximization” 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
Adarsh Kappiyath, Silpa Vadakkeeveetil Sreelatha, Xiatian Zhu, Yann LeCun, Lu Yin, Amir Bar
EBReg trains an image encoder to align augmented views while using a learned score of the feature distribution to encourage spread, with a scale constraint limiting feature magnitudes. In a matched ViT-L ImageNet-1K evaluation after 100 epochs, it reaches 75.57% linear-probe accuracy and effective rank 759, versus SIGReg’s 75.32% and 648; this comparison does not isolate the score’s causal effect.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.