A Stochastic Newton Algorithm for Distributed Convex Optimization

Explore a 35-minute lecture on a novel stochastic Newton algorithm for distributed convex optimization presented by Brian Bullins from Purdue University at the Simons Institute. Delve into the proposed method for homogeneous distributed stochastic convex optimization, where machines calculate stochastic gradients and Hessian-vector products of the same population objective. Learn how this algorithm reduces communication rounds without compromising performance, particularly for quasi-self-concordant objectives like logistic regression. Examine the convergence guarantees and empirical evidence supporting the effectiveness of this approach in optimization and algorithm design.