Dynamic Loss Function Tuning via Meta-Gradient Search

Abstract

Since they guide the optimization by measuring the difference of predictions from actual outcomes, loss functions are fundamental to the learning process of machine learning models. Historically, these roles are predefined and unchangeable during training, therefore restricting the ability of a model to adapt to changing data dynamics or learning phases. By providing a dynamic approach adjusting the loss function during training using a meta-gradient search technique this work reduces that limitation. Our method uses meta-gradients to dynamically change the parameters of the loss function in actual time, therefore matching the performance of the model. The basic idea is to improve not just the model but also the goal it absorbs, therefore providing a more flexible and tailored learning environment. We define the meta-gradient approach, in which changes to the loss function are evaluated by a superior optimization loop on next model updates. Experiments spanning many benchmarks including image classification and sequence prediction tasks show that our dynamic loss tuning produces quicker convergence, improved generalization, and higher robustness to noisy data. In many situations, the models using this adaptive approach outperform those using fixed, manually generated loss functions. This work highlights the importance of reconsidering a fundamental component of ML & offers a feasible path for automated, context-sensitive development. Giving models the ability to learn helps to create truly self-adjusting AI systems capable of independently addressing the latest challenges