Robust Implicit Backpropagation

TL;DR

This paper introduces a novel layer-wise approximation of Implicit Stochastic Gradient Descent for neural networks, enhancing robustness to learning rate sensitivity and outperforming standard backpropagation in various tasks.

Contribution

It is the first to apply ISGD to neural networks with a practical layer-wise approximation, improving training stability and performance.

Findings

More robust to high learning rates

Outperforms standard backpropagation

Applicable to various neural network tasks

Abstract

Arguably the biggest challenge in applying neural networks is tuning the hyperparameters, in particular the learning rate. The sensitivity to the learning rate is due to the reliance on backpropagation to train the network. In this paper we present the first application of Implicit Stochastic Gradient Descent (ISGD) to train neural networks, a method known in convex optimization to be unconditionally stable and robust to the learning rate. Our key contribution is a novel layer-wise approximation of ISGD which makes its updates tractable for neural networks. Experiments show that our method is more robust to high learning rates and generally outperforms standard backpropagation on a variety of tasks.