A Compression Architecture Based on Model-Code Separation1.5.2017
Compression technologies for video, images, audio, and other forms of data play a critical (and often embedded) role in today’s ubiquitous modern digital systems, software, and infrastruture. However, new compression technologies and codecs are required to meet the needs of emerging applications and platforms. In particular, there is a growing need a new generation of compression codecs that: 1) allow the traditional standardization bottleneck in compression development to be circumvented, to accelerate technology evolution; 2) allow for the rapid development of domain-specific compression technology for diverse, emerging communities; 3) support compression in the encrypted domain without requiring access to the key; and 4) are well matched to implementation on low-power mobile devices and distributed sensors. To meet these requirements, we have developed a new, practical, and general compression architecture that does not require the encoder to know the source model to achieve compression arbitrarily close to fundamental limits dictated by the entropy rate. The encoder is extremely simple to implement, and is combined with an novel inferential decoder that uses a fast, efficient message-passing algorithm in the form of belief-propagation to reconstruct the source from the compressed bits and the source model. Beyond its practical benefits, this work emphasizes the increasingly important connections between compression, statistical inference, and learning.