TL;DR: Building automatically the labeled training set with active learning for CNN. The criterion is developed on a variational inference for NN and a kronecker approximation of Fisher matrices for CNN
Abstract: One main concern of the deep learning community is to increase the capacity of representation of deep networks by increasing their depth. This requires to scale up the size of the training database accordingly. Indeed a major intuition lies in the fact that the depth of the network and the size of the training set are strongly correlated. However recent works tend to show that deep learning may be handled with smaller dataset as long as the training samples are carefully selected (let us mention for instance curriculum learning). In this context we introduce a scalable and efficient active learning method that can be applied to most neural networks, especially Convolutional Neural Networks (CNN). To the best of our knowledge, this paper is the first of its kind to design an active learning selection scheme based on a variational inference for neural networks. We also deduced a formulation of the posterior and prior distributions of the weights using statistical knowledge on the Maximum Likelihood Estimator. We describe our strategy to come up with our active learning criterion. We assess its consistency by checking the accuracy obtained by successive active learning steps on two benchmark datasets MNIST and USPS. We also demonstrate its scalability towards increasing training set size.
Keywords: Deep learning, Supervised Learning, Optimization