Improved Training of Deep Convolutional Networks via Minimum-Variance Regularized Adaptive Sampling
Abstract Fostered by technological and theoretical developments, deep neural networks have achieved great success in many applications, but their training by means of mini-batch stochastic gradient descent (SGD) can be very costly due to the possibly tens of millions of parameters to be optimized and the large amounts of training examples that must be processed. Said computational cost is exacerbated by the inefficiency of the uniform sampling method typically used by SGD to form the training mini-batches: since not all training examples are equally relevant for training, sampling these under a uniform distribution is far from optimal. A better strategy is to form the mini-batches by sampling the training examples under a distribution where the probability of being selected is proportional to the relevance of each individual example. This can be achieved through Importance Sampling (IS), which also achieves the minimization of the gradients’ variance w.r.t. the network parameters, further improving convergence. In this paper, an IS-based adaptive sampling method is studied that exploits side information to construct the required probability distribution. Said method is modified to enable its application to deep neural networks, and the improved method is dubbed Regularized Adaptive Sampling (RAS). Experimental comparison (using deep convolutional networks for classification of the MNIST and CIFAR-10 datasets) of RAS against SGD and against another sampling method in the state of the art, shows that RAS achieves relative improvements of the training process, without incurring significant overhead or affecting the accuracy of the networks.