Keywords: Cross-entropy loss, Binary classification, Low-rank features, Adversarial examples, Differential training
TL;DR: We show that minimizing the cross-entropy loss by using a gradient method could lead to a very poor margin if the features of the dataset lie on a low-dimensional subspace.
Abstract: Neural networks could misclassify inputs that are slightly different from their training data, which indicates a small margin between their decision boundaries and the training dataset. In this work, we study the binary classification of linearly separable datasets and show that linear classifiers could also have decision boundaries that lie close to their training dataset if cross-entropy loss is used for training. In particular, we show that if the features of the training dataset lie in a low-dimensional affine subspace and the cross-entropy loss is minimized by using a gradient method, the margin between the training points and the decision boundary could be much smaller than the optimal value. This result is contrary to the conclusions of recent related works such as (Soudry et al., 2018), and we identify the reason for this contradiction. In order to improve the margin, we introduce differential training, which is a training paradigm that uses a loss function defined on pairs of points from each class. We show that the decision boundary of a linear classifier trained with differential training indeed achieves the maximum margin. The results reveal the use of cross-entropy loss as one of the hidden culprits of adversarial examples and introduces a new direction to make neural networks robust against them.