Using a model of human visual perception to improve deep learning

Abstract

Deep learning algorithms achieve human-level (or better) performance on many tasks, but there still remain situations where humans learn better or faster. With regard to classification of images, we argue that some of those situations are because the human visual system represents information in a format that promotes good training and classification. To demonstrate this idea, we show how occluding objects can impair performance of a deep learning system that is trained to classify digits in the MNIST database. We describe a human inspired segmentation and interpolation algorithm that attempts to reconstruct occluded parts of an image, and we show that using this reconstruction algorithm to pre-process occluded images promotes training and classification performance.