Be Careful What You Backpropagate: A Case For Linear Output Activations & Gradient Boosting
This addresses a fundamental issue in neural network training for classification tasks, offering a practical improvement that reduces training time and enhances performance, though it is incremental as it builds on existing activation function analysis.
The paper demonstrates that saturating output activations like softmax hinder learning in classification tasks, revealing that the benefit of softmax comes from exponential gradient boosting rather than normalization, which actually worsens performance. It shows faster convergence and improved generalization on CIFAR-10, ImageNet, and PASCAL VOC 2012, with a 33% faster convergence and slightly better performance in semantic segmentation.
In this work, we show that saturating output activation functions, such as the softmax, impede learning on a number of standard classification tasks. Moreover, we present results showing that the utility of softmax does not stem from the normalization, as some have speculated. In fact, the normalization makes things worse. Rather, the advantage is in the exponentiation of error gradients. This exponential gradient boosting is shown to speed up convergence and improve generalization. To this end, we demonstrate faster convergence and better performance on diverse classification tasks: image classification using CIFAR-10 and ImageNet, and semantic segmentation using PASCAL VOC 2012. In the latter case, using the state-of-the-art neural network architecture, the model converged 33% faster with our method (roughly two days of training less) than with the standard softmax activation, and with a slightly better performance to boot.