Training robust neural networks via smooth activation functions

Abstract

Generally, the present disclosure is directed to the training of robust neural network models by using smooth activation functions. Systems and methods according to the present disclosure may generate and/or train neural network models with improved robustness without incurring a substantial accuracy penalty and/or increased computational cost, or without any such penalty at all. For instance, in some examples, the accuracy may improve. A smooth activation function may replace an original activation function in a machine-learned model when backpropagating a loss function through the model. Optionally, one activation function may be used in the model at inference time, and a replacement activation function may be used when backpropagating a loss function through the model. The replacement activation function may be used to update learnable parameters of the model and/or to generate adversarial examples for training the model.