Deep Predictive Coding Networks for Video Prediction and Unsupervised\n Learning

Abstract

While great strides have been made in using deep learning algorithms to solve\nsupervised learning tasks, the problem of unsupervised learning - leveraging\nunlabeled examples to learn about the structure of a domain - remains a\ndifficult unsolved challenge. Here, we explore prediction of future frames in a\nvideo sequence as an unsupervised learning rule for learning about the\nstructure of the visual world. We describe a predictive neural network\n("PredNet") architecture that is inspired by the concept of "predictive coding"\nfrom the neuroscience literature. These networks learn to predict future frames\nin a video sequence, with each layer in the network making local predictions\nand only forwarding deviations from those predictions to subsequent network\nlayers. We show that these networks are able to robustly learn to predict the\nmovement of synthetic (rendered) objects, and that in doing so, the networks\nlearn internal representations that are useful for decoding latent object\nparameters (e.g. pose) that support object recognition with fewer training\nviews. We also show that these networks can scale to complex natural image\nstreams (car-mounted camera videos), capturing key aspects of both egocentric\nmovement and the movement of objects in the visual scene, and the\nrepresentation learned in this setting is useful for estimating the steering\nangle. Altogether, these results suggest that prediction represents a powerful\nframework for unsupervised learning, allowing for implicit learning of object\nand scene structure.\n