Two Losses Are Better Than One: Faster Optimization Using a Cheaper Proxy

TL;DR

This paper introduces an optimization algorithm that uses a proxy function to efficiently minimize objectives with difficult-to-compute gradients, improving sample efficiency in machine learning applications.

Contribution

The paper proposes a novel algorithm combining proximal point iterations on a proxy with few stochastic gradients, achieving convergence rates comparable to SGD on smooth objectives.

Findings

Guarantees convergence at SGD-like rates for proxy-based optimization.

Enhances sample efficiency in machine learning tasks.

Applicable to synthetic data, simulators, and mixed data sources.

Abstract

We present an algorithm for minimizing an objective with hard-to-compute gradients by using a related, easier-to-access function as a proxy. Our algorithm is based on approximate proximal point iterations on the proxy combined with relatively few stochastic gradients from the objective. When the difference between the objective and the proxy is $\delta$-smooth, our algorithm guarantees convergence at a rate matching stochastic gradient descent on a $\delta$-smooth objective, which can lead to substantially better sample efficiency. Our algorithm has many potential applications in machine learning, and provides a principled means of leveraging synthetic data, physics simulators, mixed public and private data, and more.