Learning Robust Feature Representations in Deep Networks for Image Classification

Abstract

Deep learning has emerged as the method of choice for many computer vision applications. Training deep networks involves the utilization of a loss function, such as cross entropy. In this paper, we propose a novel auxiliary loss function, the Silhouette Loss, for training deep networks with the objective of obtaining feature representations that are both tightly clustered and highly separable. We are motivated by the need for well-clustered features that can generalize effectively for the classification of diverse test samples. We also introduce an adaptive scaling scheme for the regularization parameter of the auxiliary loss, which improves robustness and eliminates the selection of another hyperparameter. By training a small network with our auxiliary loss we achieve classification performance that is comparable to that of larger networks, yet our network is more efficient and utilizes much fewer parameters.