Reproduce “Don't Stop Me Yet: Sampling Loss Minima via Dissipative Riemannian Mechanics” 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
Albert Kjøller Jacobsen, Leo Uhre Jakobsen, Johanna Marie Gegenfurtner, Georgios Arvanitidis
DIMS explores alternative neural-network solutions by moving particles through the loss landscape with a loss-directed pull and speed-dependent friction. In a seven-point regression example, its samples interpolate the training data. For a LeNet trained on MNIST, DIMS reaches EMNIST OOD-detection AUROC 0.961 ± 0.0002, versus 0.846 ± 0.004 for RLA. This is one pairing; convergence is conditional, with a local-minimum guarantee.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.