EXACT: How to Train Your Accuracy

TL;DR

This paper introduces a novel optimization framework that directly maximizes expected accuracy by incorporating stochasticity into model outputs, offering a promising alternative to traditional surrogate loss functions in classification tasks.

Contribution

It presents a new method for directly optimizing accuracy through stochastic output models, addressing limitations of existing surrogate loss approaches.

Findings

Outperforms traditional loss functions in experiments

Effective for both linear and deep models

Provides a new perspective on accuracy optimization

Abstract

Classification tasks are usually evaluated in terms of accuracy. However, accuracy is discontinuous and cannot be directly optimized using gradient ascent. Popular methods minimize cross-entropy, hinge loss, or other surrogate losses, which can lead to suboptimal results. In this paper, we propose a new optimization framework by introducing stochasticity to a model's output and optimizing expected accuracy, i.e. accuracy of the stochastic model. Extensive experiments on linear models and deep image classification show that the proposed optimization method is a powerful alternative to widely used classification losses.