Reproduce “KOALA++: Efficient Kalman-Based Optimization with Gradient-Covariance Products” 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
Zixuan Xia, Aram Davtyan, Paolo Favaro
KOALA++ updates neural-network parameters using a compact directional projection of parameter uncertainty, rather than storing a full covariance matrix. It recursively approximates the needed covariance product with a minimum-norm estimate and adds a symmetry constraint. On CIFAR-100 with ResNet-50 after 100 StepLR epochs, it records 22.24% Top-1 error versus AdaBelief’s 23.07%; this result is task-specific, and positive-semidefiniteness is not guaranteed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.