Video Compression with a Predictive Neural Network

抄録

Predictive networks are a type of generative neural network model that learns to minimize the error between predicted data and real input. Prediction is used as a way to perform unsupervised learning of latent structure in the data, for example shapes and linear transformations in images. As a result, video-trained predictive networks can produce output by processing input through intrinsically stored invariances. In this study we propose to use such learned invariances as a compresssion/decompression engine for videos on spatial and temporal dimensions.