Robert C Berwick, PhD - US grants

The Grand Challenge Application Groups competition provides one mechanism for the support of multidisciplinary teams of scientists and engineers to meet the goals of the High Performance Computing and Communications (HPCC) Initiative in Fiscal Year 1992. The ideal proposal provided not only excellence in science: focussed problem with potential for substantial impact in a critical area of science and engineering) but also significant interactions between scientific and computational activities, usually involving mathematical, computer or computational scientists, that would have impact in high-performance computational activity beyond the specific scientific or engineering problem area(s) or discipline being studied. In the award to Berwick, Bizzi, Bulthoff, Jordan, Wexler, Poggio, Rivest, Winston, and Yang at MIT, the research project - High Performance Computing for Learning - has been designed explicitly to push the High Performance Computing algorithmic and architectural envelope via a CM-5 and VLSI testbed and to address many of the HPCC goals. It will advance new algorithms and software for a broad class of optimization and learning problems, tested on and directly driving operating system and architectural changes on the CM-5 (working with one of the CM-5's key architects). The learning problems addressed are essentially an entire class of modeling/optimization problems that intersect with nearly all HPCC Grand Challenge Problems.

Linguistic theory and computation have both made substantial progress over the past decade. However, the connections between the two have not been as carefully elucidated, so that the synergy between linguistic science, which in effect describes data structures, and computer science, which describes the algorithms engaging those data structures, has not been fully realized. Partly for this reason, current application systems by some estimates lag behind linguistic theory by 5-10 years. This 2-day workshop aims to bring together researchers in both linguistic science and computation to address this gap, with the aim of producing a book of collected consensus and divergent views on this important topic.//

This work will exploit learning techniques recently developed to work towards computers that learn to detect and recognize people, estimate user's gestures and communicate visually with them via a photorealistic computer-generated human face. In particular, the plan is to use tow main theoretical and algorithmic approaches to learning: Support Vector Machines and Hidden Markov Models. With these tools, two key aspects of a trainable man-machine interface will be developed: An analysis module that can be trained to estimate in real time facial expressions of the user and associated physical parameters and a synthesis module that can be trained to generate image sequences of a real human face synchronized to a text-to-speech system. The significance of the work is three-fold: (1) The project will contribute to the development of a new generation of computer interfaces more user-friendly and human-centered than today's interfaces. Such interfaces will be of direct use in education and as components of prostheses for the disadvantaged; (2) The project will integrate recently developed learning techniques to real time vision and graphics applications; and (3) The project will explore the boundaries of what is possible to achieve using 2D representations of faces rather than the more common, physically-based, 3D-based models.

9873451
Mitter
Our approach to building intelligent systems is based on the theoretical ideas of Bayesian statistical signal analysis and control-theoretic motion guidance. These theories have achieved major successes in many widely used systems. However, it has remained a challenge to create systems that integrate the analysis of low-level concrete sensor data and motor control with the use of high-level, more abstract representations of perception and action. The investigators on our team propose to attack this central issue with a variety of approaches. We believe that a very promising tool is the systematic study of hierarchical compositional data structures, in which larger scale more abstract objects are built up in stages from small-scale concrete objects. This compositional approach applies to making decisions as well as to perception: subtasks may be composed into larger tasks and reasoned with as units. Doing reasoning in such hybrid systems, it is essential that ambiguity and multiple alternatives be maintained as long as possible. For example, this means not choosing between inconsistent high-level interpretations of a signal until larger scale context is available; or replacing a local 'greedy algorithm' decision by a dynamic programming-style list of optimal conditional decisions. We discuss the role of learning in such hierarchical layered systems. Understanding, the interaction between information and control in such layered systems is a critical research issue

We plan to concentrate our research on the following aspects of layered intelligent systems:
(i) the role of the compositional and hierarchical approach in language and speech recognition; (ii) hierarchical decision making in layered systems; (iii) learning of hierarchical concepts; and (iv) interaction information and control in distributed systems. We emphasize that these topics are highly inter-related.
***

Bayesian Learning at the Syntax-Semantics Interface
Abstract

Children easily learn features of novel verbs from small numbers of scene-utterance pairs. For example, after encountering a few examples of "breaking" an object, they infer that break might require an object, e.g., John broke the glass. They also learn semantic properties. Children and adults can then generalize to other scene instances representing break. This project hypothesizes that children combine syntactic and semantic evidence to learn verb features, using a probabilistic method called Bayesian inference.

The project's first goal is to implement a computational model that can induce probability distributions on features from a very small number of scene-utterance pairs. This model will make explicit all the information sources used. Second, the project will confirm which cues are actually used by human learners in certain settings. The experimental method matches the computer model's predictions empirically, by presenting adult and child learners with training sequences of novel verbs used across varying syntactic and semantic feature situations. This project's results will advance adaptable computer systems and information-filtering, both in terms of robustness to noise and an ability to learn from a small number of examples. These results will improve the construction of a key component of natural language processing engines: the dictionary.

Input-constrained children, including the blind and deaf, still learn language, and this process continues beyond the age of five. Children are also remarkably creative in how they learn to read and write. What accounts for this robustness? For more than forty years, two distinct views have dominated the field. The first emphasizes that infant and young-child learning is qualitatively different from other novice learning, i.e., it is driven by maturation or changing brain organization. The second view emphasizes the syntax-semantics link in the acquisition of vocabulary, which seems to indicate that there is also learning machinery that is information-driven in the sense that anyone (young, old, mentally handicapped, environmentally deprived) who learns a language has to acquire elements in a certain order, e.g., "doggie" before "jump" and "jump" before "think". More recently, work in genetics and the biology of language has begun to illuminate the distinction as well as the synergy between these two views. The goal of this three-day workshop, Rich Languages from Poor Inputs, is to bridge the gap between recent theory and educational practice. It will bring together language scientists who probe these recent theoretical developments with educational practitioners so as to see how these results might inform classroom activities.

This workshop will directly inform educators as to how they can develop effective new techniques for teaching reading, writing, and interacting with the profoundly impaired. It will illustrate how educators can incorporate research on language development for children who can not learn to read easily, taking our understanding of language science into practical classroom activities. The goal is to build on children's natural talents with language, despite children's differences in abilities. This is a matter of prime importance given society's emphasis on language and reading skill attainment for all segments of the population--a true sense of no child left behind.