Computational Mathematics and Scientific Computing Seminar

Time and Location:

Speaker:

Link:

I will dive into the theory of adaptive learning rates for non-smooth convex optimization, decomposing the problem into two parts: a) estimation of the key problem unknown parameters b) Understanding when and why learning rates should be decreased. Recent developments in (a) have led to new methods that estimate the distance-to-solution, having both excellent theoretical and practical properties. For (b) we view decreasing rates through the lens of last-iterate convergence, a viewpoint that explains many practical observations that were not previously well explained, and leds to new any-time optimization methods.