On the Robustness of Monte Carlo Dropout Trained with Noisy Labels

The memorization effect of deep learning hinders its performance to effectively generalize on test set when learning with noisy labels. Prior study has discovered that epistemic uncertainty techniques are robust when trained with noisy labels compared with neural networks without uncertainty estimation. They obtain prolonged memorization effect and better generalization performance under the adversarial setting of noisy labels. Due to its superior performance amongst other selected epistemic uncertainty methods under noisy labels, we focus on Monte Carlo Dropout (MCDropout) and investigate why it is robust when trained with noisy labels. Through empirical studies on datasets MNIST, CIFAR-10, Animal-10n, we deep dive into three aspects of MCDropout under noisy label setting: 1. efficacy: understanding the learning behavior and test accuracy of MCDropout when training set contains artificially generated or naturally embedded label noise; 2. representation volatility: studying the responsiveness of neurons by examining the mean and standard deviation on each neuron's activation; 3. network sparsity: investigating the network support of MCDropout in comparison with deterministic neural networks. Our findings suggest that MCDropout further sparsifies and regularizes the deterministic neural networks and thus provides higher robustness against noisy labels.