Active Learning for Artificial Neural Network Models

Abstract

Active learning is the field of choosing informative data to train machine learning models. This thesis covers eight separate substudies investigating how to maximize the test accuracy for deep feed-forward artificial neural network models using active learning. When performing active learning, a query strategy must be specified which is why four different query strategies were examined, namely the margin, entropy, least confident, and one suggested by the authors, the least squares. The data sets MNIST, Fashion-MNIST, and CIFAR-10 were used to see how the results generalize between data sets. With the eight substudies examined, we have concluded some suggestions that should be considered to improve the test accuracy: (1) Among the query strategies examined, the margin query strategy consistently selected data which gave rise to the highest test accuracy. (2) The cumulative training method is most suitable to train feed-forward neural networks when using a query strategy. This means that the networks should be reset and retrained using all labeled data. (3) For improved performance, a query strategy should be used after the network has trained on some initially randomly selected data. (4) If the mean margin informativeness measure, used internally by the margin query strategies, starts to decrease during training, then one should consider gathering more unlabeled data or stop labeling to reduce cost. (5) The semi-supervised pseudo-label algorithm may be used to further increase test accuracy by utilizing the unlabeled data set. (6) To estimate the performance of a network without the presence of a dedicated labeled test set, one can use the randomly sampled data from (3) to create an upper and lower estimate of the test accuracy. We have shown, through empirical studies, that steps (1)-(6) are all associated with some benefit when performing active learning.