An analysis of the robustness of deep face recognition networks to noisy training labels

In recent years, state-of-the-art face recognition performance has improved by using deep convolutional neural networks. One disadvantage of these methods is their need for very large, labeled training datasets as collecting and labeling them can be time consuming and prone to error. In this work we examine the robustness of a convolutional neural network to limited training data and training data with noisy labels. We train face recognition networks with varying training set sizes and varying amounts of mislabeled samples. Our experiments show data fidelity is significantly more important than training set size; decreasing the percentage of correctly labeled samples by ten is approximately equivalent to halving the number of training samples.