- Abstract: We propose Adaptive Sample-space & Adaptive Probability (ASAP) coding, an efficient neural-network based method for lossy data compression. Our ASAP coding distinguishes itself from the conventional method based on adaptive arithmetic coding in that it models the probability distribution for the quantization process in such a way that one can conduct back-propagation for the quantization width that determines the support of the distribution. Our ASAP also trains the model with a novel, hyper-parameter free multiplicative loss for the rate-distortion tradeoff. With our ASAP encoder, we are able to compress the image files in the Kodak dataset to as low as one fifth the size of the JPEG-compressed image without compromising their visual quality, and achieved the state-of-the-art result in terms of MS-SSIM based rate-distortion tradeoff.
- Keywords: Data compression, Image compression, Deep Learning, Convolutional neural networks