Beyond Gradient Descent for Regularized Segmentation Losses

TL;DR

This paper explores alternative optimization methods to gradient descent for regularized segmentation losses, demonstrating that methods like ADM outperform GD in weakly-supervised CNN segmentation tasks, and emphasizing the importance of optimization choices.

Contribution

It introduces a new loss function inspired by MRF/CRF regularization models and shows that alternative optimizers can achieve state-of-the-art results where GD fails.

Findings

ADM outperforms GD in weakly-supervised segmentation

Smoother loss tuning improves GD results

Network design should consider optimization methods

Abstract

The simplicity of gradient descent (GD) made it the default method for training ever-deeper and complex neural networks. Both loss functions and architectures are often explicitly tuned to be amenable to this basic local optimization. In the context of weakly-supervised CNN segmentation, we demonstrate a well-motivated loss function where an alternative optimizer (ADM) achieves the state-of-the-art while GD performs poorly. Interestingly, GD obtains its best result for a "smoother" tuning of the loss function. The results are consistent across different network architectures. Our loss is motivated by well-understood MRF/CRF regularization models in "shallow" segmentation and their known global solvers. Our work suggests that network design/training should pay more attention to optimization methods.