Reproduce “Energy-based Hopfield Boosting for Out-of-Distribution Detection” 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
Claus Hofmann, Simon Schmid, Bernhard Lehner, Daniel Klotz, Sepp Hochreiter
From Johannes Kepler University Linz, the Hopfield Boosting method integrates Modern Hopfield Networks and an adaptive boosting framework into an outlier exposure paradigm for out-of-distribution (OOD) detection. This approach achieved state-of-the-art performance across CIFAR-10, CIFAR-100, and ImageNet-1K benchmarks, lowering the mean False Positive Rate at 95% True Positives (FPR95) to 0.92% on CIFAR-10.
Key insights
- Modern Hopfield Network energy functions can be effectively integrated into a training-time loss for OOD detection, moving beyond post-hoc applications.
- An adaptive boosting mechanism, focusing on auxiliary outliers difficult for the model to distinguish, significantly sharpens the in-distribution manifold's decision boundary.
- The introduced E_b energy function has a probabilistic interpretation, linking its minimization to reducing uncertainty at the ID-OOD boundary.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.