|
IAPR/ICDAR Outstanding Achievements Award Speech
Probabilistic Graphical Models in Machine Learning (Slides) (Video / Fast Download)
Sargur N. Srihari
SUNY Distinguished Professor
University at Buffalo, The State University of New York
Abstract: Machine learning is the design of computer programs that learn from data for tasks such as classification and regression. The frequently changing nature and availability of large quantities of training data makes the approach more attractive than manual recoding. Probabilistic graphical models are expressive methods for representing underlying distributions. They allow declarative representation of variable relationships and aid in understanding differences between competing approaches. This talk will describe major themes in probabilistic graphical model research giving examples of Bayesian network and Markov random field inference for pattern recognition and search. Important variations such as dynamic Bayesian networks and conditional random fields will be described.
Bio: Dr. Sargur (Hari) Srihari is a Professor in the Departments of Computer Science and Engineering and of Biostatistics at the University at Buffalo. His research led to the formation of CEDAR, the Center of Excellence for Document Analysis and Recognition, which developed the first systems for handwritten mail in the United States, handwritten addresses on US tax forms, and forensic handwriting comparison. Srihari has served as principal adviser to 35 doctoral students and hundreds of Masters students many of whom have gone on to distinguished careers in industry and academia throughout the world. Srihari has served as a member of several United States national committees, including the Board of Scientific Counselors of the National Library of Medicine and the National Academy of Sciences Committee on Identifying the Needs of the Forensic Science Community. Srihari's honors include: Fellow of the IEEE and IAPR, distinguished alumnus of the Ohio State University College of Engineering and the IAPR-ICDAR outstanding contributions award. His current research interests are in machine learning and in computational forensics.
IAPR Distinguished Speaker (Keynote Speech 1)
Document Recognition without Strong Models (Slides) (Video / Fast Download )
Henry S. Baird
Lehigh University, USA
Abstract: Can a high-performance document image recognition system be built without detailed knowledge of the application? We have benefited from the statistical machine-learning revolution of the last twenty years, and as a result we rely less on hand-crafting special-case rules and more on learning models from labeled samples. But urgent questions remain. If we can't collect and label enough real training data, how helpful is it to complement them with data synthesized using generative models? When is it safe to rely on synthetic data? If we can't manage to craft (or train) a single complete, near-perfect "strong" model to drive recognition, can we make progress by combining several imperfect or incomplete "weak" models? Can recognition that is carried out jointly over weak models perform optimally while still running fast? Can a recognizer automatically pick a strong model of its input? Must we pre-train strong models for every kind ("style") of input expected, or can a recognizer adapt to unspecified styles? Can weak models adapt autonomously, without any human intervention, growing stronger and so driving performance higher monotonically? Can one weak model "criticize"---and then correct---other weak models, even while it is being criticized by them? After fifteen years of research, we now have partial answers to these questions, and successful applications to show. I'll illustrate the evolution of the state of the art with concrete examples, and point out open problems.
(Based on work by and with T. Pavlidis, T. K. Ho, D. Ittner, K. Thompson, G. Nagy, R. Haralick, T. Hong, T. Kanungo, P. Chou, G. Kopec, D. Bloomberg, A. Popat, T. Breuel, E. Barney Smith, P. Sarkar, H. Veeramachaneni, J. Nonnemaker, and P. Xiu.)
Bio: Dr. Henry S. Baird is Professor of Computer Science & Engineering at Lehigh University. Prior to joining academia he was a manager of research at Bell Labs and the Palo Alto Research Center. He has been elected Fellow of the IAPR as well as the IEEE, and he has received an ICDAR Outstanding Contributions award. He served on Editorial Boards for IEEE Trans. on PAMI, Computer Vision and Image Understanding, and the Int'l J. on Document Analysis and Recognition. He has authored three books and over eighty technical articles; he holds seven patents; and he has led six conferences and workshops.
IAPR Distinguished Speaker (Keynote Speech 2)
Chinese Paleography, Calligraphy and Pattern Recognition: Styles and Scripts in Excavated Ancient Chinese Documents (Slides) (Video / Fast Download)
Wen Xing
Dartmouth College, USA
Abstract: In the past two decades, over 220,000 bamboo and wood slips of ancient Chinese documents were discovered in China. Just as the discovery of the Dead Sea Scrolls significantly changed the study of Judeo-Christian biblical tradition, these excavated documents, most of which were unavailable in the transmitted textual tradition, have dramatically changed the study and understanding of almost every aspect of ancient China. As original work of ancient Chinese calligraphy produced in different historical periods and geographic areas, the excavated documents are vivid demonstrations of structural complexity and aesthetic characteristics of ancient Chinese scripts. It is argued that an interdisciplinary study of Chinese calligraphy, paleography and pattern recognition provides an innovative digital approach to the interpretation of excavated ancient Chinese documents and the revolution of traditional Chinese paleography.
Bio: Dr. Wen Xing is Associate Professor of Asian and Middle Eastern Languages and Literatures and of Asian and Middle Eastern Studies at Dartmouth College. He is a specialist in traditional Chinese painting and calligraphy, as well as an expert of excavated ancient Chinese documents inscribed on bamboo and silk. A recipient of a number of competitive national and international awards and grants, he is the author of Research on the Silk Manuscript Zhouyi (Beijing, 1997), Written on Bamboo and Silk: Thought and Schools in Early China (Taipei, 2005), and Tombs, Texts, and Transcriptions: An Introduction to Excavated Chinese Texts (San Antonio, 2005), etc. He is also the series Editor of Excavated Chinese Classics in Translation, published by MerwinAsia in Portland, M.E.
IAPR Distinguished Speaker (Keynote Speech 3)
Marc Wilhelm Küster
University of Applied Sciences Worms
Abstract: The lead medium of the humanities is text, but text with special characteristics that can be quite different from a normal monolingual article in most modern scripts. Text that can be derived from manuscripts, from retrodigitization of previous scholarly publications such as critical editions and dictionaries, from books printed centuries ago, applying conventions no longer in force today.
The keynote identifies four major challenges for recognizing humanities data: Unusual characters, unusual layouts, unusual semantics and unusual segmentations. Each challenge is illustrated with concrete examples taken from a variety of times and places, starting with cuneiform tablets, an extract from a Greek manuscript, a page from a multilingual critical edition, a renaissance print, a lemma from a scholarly dictionary, and some more.
In addition, scholarly humanities data is typically marked up using domain-specific rich XML-based formats based on the TEI P5 guidelines. Any format that an OCR program produces must be sufficiently rich to permit for a mapping on TEI-compliant markup in order to be capable of reproducing the full richness of the original.
A closer view at the TextGrid virtual research environment for the humanities and its Text-Image Link Editor (TBLE) demonstrates how scholars currently tackle these tasks. It analyzes where automatization can facilitate their task and enable new dimensions of research.
Bio: Marc Küster got the Diplom thesis in physics on gas-surface-interactions at the University Osnabrück in 1994, the Master in philology and history on the French literary journal Divan at the University Osnabrück in 1997. During 1997-2001, he worked at the University Tübingen's Computing Centre in the department of Literary and Documentary Data Processing. During 2001-2005, he was a Co-founder and co-director of the XML service provider Saphor GmbH, and worked on his dissertation on the tradition of alphabetic ordering from cuneiform to computers ("Geordnetes Weltbild") at the Faculty of Modern Philology at the University Tübingen. Since 2005, he has been a Professor for Web Services and XML Technologies at the Faculty for Computer Science and Telecommunications at the University of Applied Sciences Worms. His research interests include Web Services, XML technologies, Semantic Web, and the application of information technology in the humanities (eHumanities/Digital Humanities).
|
|
