Remarks at Gary Kopec's Memorial
I've been asked to say a few words about Gary's professional life, from the viewpoint of a colleague and a friend.
Gary and Marcia's first formal contact with Xerox was in 1981, when they were offered jobs at PARC by Dick Lyon. They decided instead to accept offers from the Schlumberger research lab, also in Palo Alto. When they arrived there, they were surprised to find that their new boss would be -- Dick Lyon!
Gary and Marcia finally arrived at Xerox, together, in 1988, and in the following years both Gary and Marcia have been, at different times, my first-level managers. I could not have been more lucky. When Gary arrived, the plan was for him to manage an eclectic and essentially unmanageable group of individuals. It was a thankless job, but Gary gamely took it on. Over the next ten years, Gary had several opportunities to recruit people to work with him at Xerox, and he patiently staffed up a group of colleagues, all with a mathematical and statistical bent, to work on document image analysis.
Gary's most recent vita lists 12 issued patents and 68 publications. Most of his patents are related to the field he founded with Phil Chou - document image decoding; other patents will issue later. His prodigious output of publications spans signal processing and recognition, for both speech and images. However, not all of his work has been published. For example, Gary received an infamous rejection letter when he submitted his 1992 Xerox technical report on ``Row-major scheduling of Image Decoders". The editor could not find any reviewers for the paper! There were several reasons, but perhaps the most important is that such rejections are an occupational hazard for pioneers. Gary lived on the edge. He understood and used all the techniques that were available and required, from formal grammars to statistical pattern recognition, machine learning, and signal processing. And he needed them all to move from 1-dimensional hidden markov models used in speech recognition to formal methods for solving 2-dimensional image problems.
Gary was also a consummate technician. Whereas a program to decode a line of text is not too difficult to write, a decoder for a musical score can be a hundred times more complicated. How was Gary to write such a program? It would have been too painful to do it by hand, so instead he wrote a lisp program to write the decoder. Programs that write programs that write programs. He understood the subtleties of ``going meta." Graphs, parsers, grammars ... the stuff of discrete mathematics and computer science. Sampled data, stochastic models, nonlinear optimization, maximum likelihood ... the stuff of electrical engineering. Gary needed it all, and he did it all. He was a Renaissance man, and he fit perfectly into the digital processing renaissance of the last 25 years. He went ``meta" at low levels, too. The lisp code was too slow for the inner loops that did the image convolutions, so he wrote the convolution code in C. It was still slow, because of the immense amount of computation required. So he wrote a lisp program that would write a C program, with the specific template convolutions hard-coded. Loops were all unrolled; computation was minimized. After compilation, it ran about 5 times faster than the generic C version.
Everyone who came in contact with Gary had the opportunity to learn. Gary enjoyed sharing his thoughts with inquiring minds, and he had a sequence of students that worked with him on image decoding. Several of these students - Anthony Kam, Mauricio Lomelin and Jesse Hull - were MIT students that did masters theses under Gary's supervision. These theses are required reading today for anyone who wants to understand the field. Gary succeeded so well with the very brightest students because he had the ability to communicate the beauty and generality and, perhaps above all, the inevitability of his methods.
I was a bit slow to grasp this inevitability, but I remember when it struck, it was like a flash of lightning. To explain why, I should go back. When Gary came on the Xerox scene in 1988, I was an old-timer, relatively speaking, but had recently switched from magnetic and optical engineering to the strange world of image analysis. At my first and only computer vision conference, I was puzzled that people in the field were complaining that so little progress had been made in the past 25 years. I had already decided that for analysis of the large images you get when you scan documents, image-based operations, as opposed to sets of rules on non-image structures, were the most productive direction to proceed. Rules just didn't work well. Gary placed it all in perspective. The same problem had occurred in speech recognition: people tried for 20 years to build rule-based systems, but progress was slow. The big break came with the introduction of hidden markov models, which had both sufficient internal complexity and could be trained automatically using input data. And speech recognition never looked back. Gary predicted that data-centric statistical models, particularly hidden markov models with an underlying grammar, would likewise eventually dominate among successful image recognition techniques. For 1989, this seemed to be a radical suggestion.
I have persisted in using ad hoc methods and parameters to get quick results, but Gary consistently followed his principled approach for image recognition: use the data to build the models for image production, and then use the models to find the most likely generating source for a particular observed image. This was elegantly described by his ``communications model" of image recognition, and his use of the term ``decoding" to describe it. Inevitability! The best models are constructed using the image data, and for any image, Gary's decoders then give you the best possible guess for the input (that is, the best recognition output), given the model.
The best is hard. But Gary would never settle for anything less than the best. I think that is the most salient characteristic of Gary's approach - to everything. His technical talks were marvels of clarity. I invariably had the feeling that it would have been impossible to improve on them. Gary thought so deeply about his work that it was easy to believe that he had everything figured out. But he was also able to formulate the outstanding problems in a way that attracted students by their practical and fundamental nature.
Gary set out his document image decoding research agenda in an internal talk in 1992. ``Radical" is a favorite word with upper management at PARC; we're constantly reminded that it comes from the Greek for ``going to the root". Well, this was as radical a research proposal as I'd ever heard. He emphasized that it was important to focus on the problem, and to find general and formal solutions, as opposed to ad-hoc ones. He said, ``It's better to do the wrong thing well," by which I believe he meant that if you have a general formal approach, your solutions will improve as you eventually fix your models. Gary demanded generality by the unification of a common architecture, not by the aggregation of ad-hoc approaches. He believed in learning from shortcomings (he called it ``ignorance modelling"), where you focus on the gap between what is known and what needs to be known - in his words: ``tailor, train, tune". And he put his methodology in stark contrast to the current ad-hoc approaches, where the systems were complex, and thus often lacking peer review and validation; they were idiosyncratic, and thus lacking points of commonality; and they were strongly self-interacting, where changes in prior knowledge have unpredictable system effects.
Gary's respect for data was unwavering, and it went beyond image decoding. For example, I was working on a fast method for detecting skew in scanned document images. Henry Baird had previously approached the problem by finding the connected components and looking for alignment of their lower edges. I'd tried using binary image operations to select instead a subset of edge pixels, from which to search for alignment, and the results were good. Gary immediately said: ``you don't know which pixels are best, so why not use all the pixels!" Use all the data. I did, and - yes - the results were even better. And this is the basis of our standard method.
Gary affected all of us in many ways. He had a gentle and thoughtful personality, and was always available to consider issues, whether technical or personal. And, when things got irrational, as they sometimes did, he was a forceful advocate for sanity.
Gary mentored everyone around him in a quiet way, and he was respected for his intellect and wisdom by his colleagues in many fields. He could just as easily have been a university professor. Xerox gave him the opportunity both to pursue his research with few distractions, and to have students and colleagues to work with. But Gary also had very strong ties to colleagues in both industry and universities. His most ambitious academic collaboration was with with the NSF-funded Berkeley Digital Library Project, in which he played a major and virtuoso role. He scanned and decoded a huge database of dams in California, that had originally been typed, and was only available as 600 pages of tables on paper. He then generated links between several different views of this data for a web browser interface. This was immediately useful to state water resource employees. And using it, in five minutes Gary could demonstrate to anyone the value of special recognizers, and how such automatically generated hyperlinked web documents made it so much easier to find information.
Gary hired Phil Chou in 1990, and one of Phil's most cherished memories was from his first day on the job. Gary and Phil had a long conversation about the technical problems of building consistent probability models for images built up of characters on a page. After a pause towards the end of the conversation, Gary nodded in satisfaction, and remarked about the prospect of their future collaboration, in his typically understated way, ``This is going to be good ..." And it was!
Beyond all his talents - his ability to find the most important problems, and his optimism and energy to attack and solve them - Gary had an inner strength that confounded me. He was probably in better condition than anyone in the research center. He exercised every day, typically running about 5 miles in all weather conditions. I never tried to run with him, because it would have been impossible to keep up. And when he got sick a year and a half ago, he approached his fate with a courage I still cannot imagine. He forewent drastic invasive procedures that might have prolonged his survival, because he believed that every day counted and the quality of his days was paramount. He retained his interest in all things, mundane and profound, up to the end.
All of us who were touched by his gifts, and by his generosity, have been enriched in special ways that we will never forget.
-- Dan S. Bloomberg, January 24, 1999
No comments:
Post a Comment