January 16, 2007
First day of classes!
January 19, 2007
-- Bill Dolan (Microsoft)
Title: "Paraphrase as an emergent property of the web."
Many natural language processing applications require the ability to recognize when two text segments ? however superficially distinct ? overlap semantically. Question-Answering (QA), Information Extraction (IE), command-and-control, and multi- document summarization are examples of applications that need precise information about the relationship between different text segments. The last few years have seen a surge in interest in modeling these phenomena, but efforts to machine-learn paraphrase recognition/ generation models have been hampered by a lack of training data. While some datasets (e.g. parallel news articles, multiple translations of the same original text) do exist, their scale is too limited to support broad-coverage semantic models. This talk will describe some of the techniques used to model paraphrase relationships, discuss the limitations stemming from current restrictions on data size, and finally, describe a web- based data collection technique that addresses many of these limitations. Bill Dolan is a Senior Researcher and the manager of the Natural Language Processing Group at Microsoft Research in Redmond, Washington. He has worked on many aspects of semantic processing, including word sense disambiguation and MindNet, a large-scale lexical knowledge base built automatically from free text. His current interests include paraphrase recognition/generation and machine translation.
January 26, 2007
February 2, 2007 (SLC)
-- Edward (Joe) Redish
Professor of Physics, University of Maryland
Title: "Generalization in Physics: Perspectives from Practice and Theory"
Children learn many basic components of their adult knowledge by forming patterns and generalizing - grammar, counting, basic math, and physical phenomenology. The question addressed in this talk is: what role does generalization play in scientists developing new knowledge and students learning existing complex science? This talk has three parts. In the first part, I will demonstrate a pedagogical example from introductory physics that shows how generalization and specialization take turns leading in an intricate, interactive dance in the construction of new scientific knowledge. In the second part, I will show how we applied a similar process of moving between generalization and specialization to understand how students approach problem solving in introductory physics. Analyzing ethnographic data of students solving physics problems, we conclude that much of their behavior can be described by a cognitive structure we refer to as an epistemic game - a local coherence in behavior. The games they choose can be either productive or counterproductive in helping them solve a problem. In the third part, I will raise the question whether the concepts of generalizing and specializing, which were useful in describing the processes in the first two parts of the talk, are the best way to describe what students are doing as they learn. A more useful language might be to talk about activation, association, binding, and contextual framing.
E. F. (Joe) Redish is a Professor of Physics and an affiliate Professor of Curriculum & Instruction at the University of Maryland. For over twenty-five years he was an active researcher in theoretical nuclear physics. He always had a strong interest in teaching, and, upon discovering that a classroom was an even more complex strongly-interacting many-body system than a nucleus, switched his field of research to physics education. For more than a decade, Joe has been a leader in helping to establish a discipline-based education research community within physics. He has researched a variety of topics ranging from the implications of student expectations for their behavior in introductory physics to the difficulties advanced students have with quantum mechanics. His current interest is in building theoretical models for science education with ties to neuroscience, cognitive science, and the behavioral sciences. He is the winner of numerous awards for his education work including the Millikan Medal from the American Association of Physics Teachers, the Director's Distinguished Teaching Scholar award from the National Science Foundation, and a Distinguished Scholar- Teacher award at the University of Maryland.
February 15, 2007 (Psychology Talk held in MUEN D-430/428)
-- Jeremy Wolfe
Professor of Ophthalmology, Harvard
Title: "Visual Search: Is it a matter of life and death?"
Like Gaul, this talk is divided into three parts: 1) I will give an introduction to the problem of visual search and to the Guided Search model that my lab has been working on for a number of years. The details of Guided Search will be discussed in my other talk. 2) I will place the problem of search into the larger context of visual perception and show how our need to use selective attention leads to some interesting perceptual errors. 3) Finally, in an effort to convince you that my particular intellectual obsessions are, in fact, "a matter of life and death". I will discuss an important practical problem in search. Rare targets are hard to find simply because they are rare. We ask people to find rare targets in some very important tasks like airport baggage screening and routine mammography so, if low target prevalence makes search difficult, this could be a real problem.
February 16, 2007 (To be held in MUEN E-214)
-- Jeremy Wolfe
Professor of Ophthalmology, Harvard
Title: "Guided Search 5.0: Toward a "scenic" upgrade"
Abstract: Guided Search is a model of human visual search behavior. It has several core beliefs and an endless collection of details and problems. Belief One: There is a serial bottleneck between the massively parallel first stages of visual processing and the massively parallel processes that permit object recognition. Only one object (or perhaps a very few objects) can be recognized at one time. Belief Two: Access to that bottleneck is under the control of visual selective attention. Attentional process select objects at a rate of something like 20-40 objects per second. Belief Three: Selection can be guided by a limited number of attributes that are abstracted from those massively parallel first stages of visual processing. Assuming that we get beyond those basics, we will discuss: Problem One: How can we account for the distributions of reaction times. Problem Two: How do we know when to quit when we don't find a target? Problem Three: Can Guided Search, a model of highly artificial laboratory search tasks, be modified to account for search in real scenes. As a hint about why this might be hard, try to count the number of objects (the "set size") in any real scene.
February 23, 2007 (To be held in E-214)
-- Christer Samuelsson
Center for Computational Learning Systems Columbia University
"Artificial Intelligence and Machine-Learning Techniques to Market Prediction and Automatic Trading"
This talk will cover how the author applied Artificial Intelligence and Machine-Learning techniques to market prediction and automatic trading. It starts off with an analysis of Renaissance Technology's modus operandi ---a company that made quite a few IBM speech scientists and computational linguists very wealthy---then delves into the differences between using HMMs for ASR and for market prediction. It shows how standard alignment techniques dazzled financial professionals by inferring trade sizes from quote sizes. The talk then covers in painful mathematical detail how a physical analogy was used to model the process where market participants search for the elusive equilibrium price, which in turn gets jerked around by incoming news. It then covers in less mathematical detail an example-based machine-learning approach to market prediction, and rounds off with a quite successful attempt at automatic trading, rather than prediction, using a reinforcement-learning scheme to find appropriate trade moves. The lessons learned during two-and-a-half years are summarized into nuggets of wisdom that may be of interest to intra- and inter-day traders alike, as well as to casual investors. Caveat Auditor: this talk will not reveal any Lehman or Rennaissance trade secrets, nor how to become a millionare through automatic trading. For the latter, you must think for yourself and exploit your own insights into economics, markets, and marketmicrostructure.
March 2, 2007 (SLC)
-- Keith Holyoak
Distinguished Professor of Psychology, UCLA
Title: "Analogy in the Mind, the Brain and the Classroom"
Reasoning by analogy is a powerful tool for human reasoning and learning. How can it be used effectively? What are the underlying cognitive mechanisms? How do these mechanisms develop in children? How are they realized in the brain? I will survey our recent work addressing these questions, using methods spanning a range from cross-national comparisons of the use of analogies in mathematics instruction, to functional neuroimaging of reasoning with simple analogy problems.
March 16, 2007
Physics Department, CU
Title: "Understanding Student Learning in Physics: The Role of Representation and Analogy"
April 13, 2007
RMPA - Battig Symposia on Memory (Denver Tech Center)
April 20, 2007
Leeds School of Business
Title: "Value Construction and Bidding Behavior in Descending and Ascending Auctions."
We report two experiments examining how a set of motivational, cognitive and situational factors drives consumers' value construction and bidding in auctions involving real products. A motivational antecedent, bidder goals (winning the item versus acquiring it at a price consistent with their value) is examined along with two cognitive factors (a) value precision (operationally, the width of a provided price range) and (b) value salience (whether or not subjects' value for the item is pre- measured). The experiments involve a descending and an ascending auction respectively, each embedding manipulations of a situational variable (deliberation time available to the bidder at each price step). The results show that in both auction formats, winning focus bidders bid higher than value focus bidders. However, the effect is attenuated when values are more precise and/ or salient. Deliberation time effects differ across auction formats. In descending auctions, a longer time elicits higher bids from winning focus bidders but not from value focus bidders. In ascending auctions, a longer time lowers bids from value focus bidders, but not from winning focus bidders. We compare these effects across the two studies to explore the behavioral underpinnings of consumer value formation and bidding in descending and ascending auctions.
April 27, 2007
ICS Member Poster Session
May 1, 2007
--Sargur N. Srihari
University of Buffalo, The State University of New York
Title: "Computer Processing of Handwriting in Documents"
Handwriting is a natural means of recording personal information and continues to be used in education. Even with the ubiquity of computers, handwriting is used in forms, in legal documents and in postal addressing. Although computer recognition of handwriting seems to be a solved problem with the ubiquity of PDAs and tablet PCs—where writing is on specialized surfaces (called dynamic handwriting)—recognition of handwriting on paper documents (or static handwriting) poses numerous challenges. The talk will describe recent advances with applicability to postal address recognition, signature/writer verification and automatic scoring of responses to prompts in school testing. A system for handwriting comparison for forensics and for searching handwritten document repositories will be demonstrated.
May 11, 2007
Mayer, Goldin-Meadow, Paul Herr