Reproduce “Continuized Acceleration for Quasar Convex Functions in Non-Convex Optimization” 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
Jun-Kun Wang, Andre Wibisono
This paper from Jun-Kun Wang and Andre Wibisono at Yale University presents an optimal accelerated algorithm for minimizing quasar convex functions, achieving an O(1/√ε) convergence rate with only a single gradient evaluation per iteration. The method, based on continuized discretization, demonstrates strong empirical performance when applied to Generalized Linear Models with common link functions like logistic and ReLU.
Key insights
- Quasar convexity, a relaxed form of convexity where the negative gradient always forms an acute angle with the direction to the minimizer, provides sufficient structure for accelerated non-convex optimization.
- The continuized discretization approach allows for designing and analyzing optimization algorithms in continuous time using differential equations, which can then be discretized to yield practical and theoretically sound algorithms.
- Objective functions for training Generalized Linear Models (GLMs) with common link functions (e.g., logistic, ReLU, quadratic) satisfy (strong) quasar convexity under mild assumptions, expanding the applicability of accelerated methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.