Year: 2015
Type: dissertation
Abstract: Data is compressible by presuming a priori knowledge known as a data model, and applying an appropriate encoding to produce a shorter description.The two aspects of compressiondata modeling and coding -however are not always conceived as distinct, nor implemented as such in compression systems, leading to difficulties of an architectural nature.For example, how would one make improvements upon a data model whose specific form has been standardized into the encoding and decoding processes?How would one design coding for new types of data such as in biology and finance, without creating a new system in each case?How would one compress data that has been encrypted when the conventional encoder requires data-in-the-clear to extract redundancy?And how would mobile acquisition devices obtain good compression with lightweight encoders?These and many other challenges can be tackled by an alternative compression architecture.This work contributes a complete "model-code separation" system architecture for compression, based on a core set of iterative message-passing algorithms over graphical models representing the modeling and coding aspects of compression.Systems following this architecture resolve the challenges posed by current systems, and stand to benefit further from future advances in the understanding of data and the algorithms that process them.I am indebted to the help of many people without whom this work could not have come to fruition.Foremost among them is my advisor Professor Gregory Wornell.When I first came to MIT, I had an inkling about doing the type of research represented by this work, but it would take a long journey of learning and reflection to get here.Without his investment in me, his tremendous patience and faith, his way of developing my skills and perspectives, his long-term support of my research goals, his interest and encouragement, or his technical input, this journey would have been unimaginable.For this support and guidance I feel extremely grateful, and for having his deep insight, broad vision, and consummate intellectual ideals to draw on I feel fortunate.Most of all, being enabled to create what I genuinely set out to create is my best dream fulfilled, which significance cannot be properly measured in words.I also owe many thanks to my thesis committee for their helpful comments.Professor Devavrat Shah raised important questions that clarified parts of the work related to coding, and Dr. Emin Martinian contributed additional perspectives that led to a more refined view on performance.Over the years I have learned from my committee in many ways.Sometimes it is through their teaching, and other times through mentorship and collaboration.The seeds of a few ideas came partly from my earlier collaboration with Dr. Martinian and from his prior work on the subject.A number of people contributed additional input into this work.Ying Liu introduced several graphical algorithms to me; Chong Wang offered a detailed explanation of Ising models; Venkat Chandar pointed out the linear processing bound due to Ancheta; David Romero referred me to the literature on large-alphabet LDPC codes; Ying Yin provided expertise on sequential hierarchical graphical models; Da Wang made suggestions about existing compression algorithms; Qing He referred me to the literature on neural networks and discriminative models; Professor Bill Freeman gave me a crash course on modern image models; and Professor Yuval Kochman, who was also a collaborator when we overlapped at SIA, identified connections to more related problems.There are others with whom I have exchanged ideas about the content or presentation of the work.I would like to thank all of them.Given the practical angle of this work, I greatly benefited from prior research conducted in the industry with these additional collaborators: Drs.Jonathan Yedidia and Ali Azarbayejani at MERL where I developed practical graphical algorithms; Drs.John Apostolopoulos, Mitchell Trott, and Ton Kalker at HP Labs where I worked on information, multimedia, and security problems; and Drs.Sanjeev Mehrotra and Jin Li at Microsoft Research where I developed error correction systems and learned about audio compression.These researchers made a significant impact on the perspectives and tools I used to approach this work.In many ways, this topic also returns to my own earlier academic experience, and I thank those who first introduced me to a rich and beautiful field and set me on the path of finding out more: I enjoyed interacting closely with Profs.
Author: Ying-zong Huang
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2154275610