Architectures for Lossy Compression: A Fundamental Study

点击翻译按钮获取中文摘要

Information sources (voice, images and video) are analog in nature: they can be represented by real numbers. On the other hand, the common currency of today's information age is digital: an abstract sequence of bits. The analog-digital separation architecture is the dominant mode of operation today, due primarily to its modularity and scalability. While still not employing joint source-channel coding, lossy compression might be helped by ``analog-tagging'' the digitally compressed information sources (example: most and least significant bits).In this project, a fundamental view of lossy compression to shed insight into what architectures are appropriate for different network configurations is taken. An important component of this study is the focus on simple statistical knowledge of the analog information source. On the reconstruction side, of particular interest is the quadratic fidelity criterion. The goals of this project are: (a) to understand regimes where the analog-digital separation architecture is optimal; (b) to identify the appropriate analog tags to digital representation that can be best utilized when it comes to communicating over a network; (c) identify scenarios where such new architectures yield significant improvement over the conventional analog-digital separation architecture. Recent progress by the principal investigator in resolving decades-long open problems in distributed compression and multiple description is used as the starting point for the research agenda.