Abstract: As we move along the scale from teleoperation towards collaboration, human-robot interactions become more complex and require that the human and the robot share more common knowledge about the world and how things within the environment are related. At the collaborative level of interaction, the robot and human must exercise mixed initiative in solving a problem, each taking advantage of their unique skills, location, and perspective of the current situation. We believe that at this level and beyond, the robot will need representations and procedures that are similar to those used by humans, in order to collaborate successfully. Our working hypothesis is that a system that uses representations and processes similar to a person's will be able to collaborate with a person better than a computational system that does not. I suggest three reasons for the representational hypothesis and then describe empirical and computational evidence in several domains.
Biographical Sketch: Alan C. Schultz is the Head of the Intelligent Systems Section, Navy Center for Applied Research in Artificial Intelligence at the Naval Research Laboratory in Washington, DC. His research is in the area of evolutionary computation, evolutionary robotics, learning in robotic systems, robot/human interfaces, and adaptive systems. He is the recipient of an Alan Berman Research Publication Award, and has published over 40 articles on machine learning and robotics. Alan is currently the co-chair of the AAAI Symposium Series, and chaired the 1999 and 2000 AAAI Mobil Robot Competition and Exhibitions.

Abstract: The talk will review concepts, principles, and mathematical tools that were found useful in applications involving causal reasoning. The principles are based on structural-model semantics, in which functional (or counterfactual) relationships, representing autonomous physical processes are the fundamental building blocks. This semantical framework, enriched with a few ideas from logic and graph theory, enables one to interpret and assess a wide variety of causal and counterfactual relationships from various combinations of data and theoretical modeling assumptions. These include:
(1) Predicting the effects of actions and policies
(2) Identifying causes of observed events
(3) Assessing direct and indirect effects,
(4) Assesseing the extent to which causal statements are corroborated by data
(5) Assessing explanations of events in a specific scenario. For background information, see Causality (Cambridge University Press, 2000), or www.cs.ucla.edu/~judea/, or the following papers:
gentle-introduction
paper1
paper2
paper3
Biographical Sketch: Judea Pearl is a professor of computer science and statistics at the University of California, Los Angeles. He joined the faculty of UCLA in 1970, where he currently directs the Cognitive Systems Laboratory and conducts research in automated reasoning, decision under uncertainty, causal modeling, and philosophy of science. He has authored three books, Heuristics (1984), Probabilistic Reasoning (1988), and Causality (2000). A member of the National Academy of Engineering and a Fellow of the IEEE and AAAI, Judea Pearl is the recipient of the IJCAI Research Excellence Award for 1999, the AAAI Classic Paper Award for 2000, the Lakatos Award for 2001, and the ACM Alan Newell Award for 2003.

Abstract: The talk addresses the problem of how to integrate into a classical logical framework transitions which change the knowledge state of an agent. The transition logic achieves this integration by defining a deductive relationship |- (T,<=) among formulas for a partially ordered set (T,<=) of transitions. This novel integration is closer to human reasoning as experienced in natural language use than previous approaches to this fundamental problem. It provides a unified framework for practical reasoning with the potential of a full exploitation of the maturing techniques from classical deduction and from special applications such as planning, explanation, diagnosis, nonmonotonic reasoning, and others. ARTICLE
SLIDES
Biographical sketch: Wolfgang Bibel is currently Professor emeritus for Intellectics at the Department of Computer Science of the Darmstadt University of Technology in Germany. He also maintains an affiliation with the University of British Columbia in Vancouver, Canada, as an Adjunct Professor. Before moving to Darmstadt in 1988 he was a Professor at the University of British Columbia in Vancouver and a Fellow of the Canadian Institute for Advanced Research. Before that he worked as a senior researcher at the Technical University in Munich and held visiting professorships at several universities (Rome, Duke, Karlsruhe, Saarbrücken, Wayne State, et al).
Wolfgang Bibel received his PhD with a thesis in Mathematical Logic in 1968 and his Diploma in Mathematics in 1964 from the Ludwig-Maximilian University in Munich. His more than 200 publications including several books cover a variety of topics in Artificial Intelligence/Intellectics such as automated deduction, architecture of deductive systems, knowledge representation and inference, planning, learning, program synthesis, aspects of applications of AI in disciplines like Psychology, Sociology, Education, and Politics.
Professor Bibel for many years served on more than a dozen boards of journals and book series, for instance as Section Editor of the Artificial Intelligence Journal. In the years 1992 through 1998 he coordinated the national program on Automated Deduction in Germany. He was heavily involved in the foundation of Artificial Intelligence research and organization in Germany and Europe.
Dr. Bibel is a fellow of the American Association for Artificial Intelligence (AAAI) as well as of ECCAI, the European AI organization which he founded, and received the Donald E. Walker Distinguished Service Award from IJCAII, the International Joint Conferences for Artificial Intelligence Inc., among other awards.

Abstract: The Spoken Language Systems Group in MIT's Computer Science and Artificial Intelligence Laboratory has been developing spoken dialogue systems for nearly 15 years. These systems typically provide access to databases containing information of interest to a traveller, such as flight status and schedules, traffic and weather reports, navigation assistance, or a hotel or restaurant guide. Through experience acquired while developing such systems, we have identified two critical research topics that we believe, if solved, would lead to significant breakthroughs in the future widespread adoption of such systems. The first topic is the development of a "generic dialogue manager," which can, ideally, be configured for a specific domain by simply providing a semantic ontology of the domain, often inferred from the database itself. A necessary component of success is the ability to configure a simulation of dialogue interaction, complete with simulated user utterances, to stress test the system prior to its release to real users. The second topic concerns the issue of adaptable and flexible vocabulary. This includes just-in-time vocabulary updates of the recognizer to reflect information being presented to the user, as well as the ability for the user to update (personalize) the system's lexicon through spoken interaction. This talk will give an overview of our research, followed by discussion of where we stand now on the two research topics identified above.

Brief Biographical Sketch:

Stephanie Seneff is a Principal Research Scientist in the Spoken Language Systems Group at the MIT Laboratory for Computer Science and Artificial Intelligence. She received the B.S. degree in Biophysics from MIT in 1968, the M.S. and E.E. degrees in Electrical Engineering and Computer Science in 1980, and the PhD degree in EECS in 1985, also from MIT. During the 1970's, she was a member of the Research Staff at MIT Lincoln Laboratory, where her research topics included speech recognition, speech synthesis, and voice encoding. Her doctoral thesis concerned a model for human auditory processing of speech. She has been conducting research at MIT for over fifteen years, during which time she has pursued a wide range of topics related to the development of spoken dialogue systems, including speech recognition, subword linguistic modelling, natural language understanding and generation, and discourse and dialogue modelling. She played a leadership role in the development of the Galaxy architecture for spoken dialogue systems, which is widely used in the research community. She has published nearly a hundred papers in conference proceedings and technical journals. She has supervised many students pursuing MEng,, M.S., and Ph.D. degrees at MIT. She is a member of the IEEE Society for Acoustics, Speech and Signal Processing, where she served on the Speech Technical Committee from 1987 to 1992. She is also a member of the Editorial Board for the Speech Communications journal, and has served as a member of the Permanent Council for the International Conference on Spoken Language Systems (ICSLP).

Abstract: Adaptive Resonance Theory (ART) neural networks model real-time prediction, search, learning, and recognition. ART networks function both as models of human cognitive information processing [1,2,3] and as neural systems for technology transfer [4]. A neural computation central to both the scientific and the technological analyses is the ART matching rule [5], which models the interaction between top-down expectation and bottom-up input, thereby creating a focus of attention which, in turn, determines the nature of coded memories.

Sites of early and ongoing transfer of ART-based technologies include industrial venues such as the Boeing Corporation [6] and government venues such as MIT Lincoln Laboratory [7]. A recent report on industrial uses of neural networks [8] states: "[The] Boeing Neural Information Retrieval System is probably still the largest-scale manufacturing application of neural networks. It uses [ART] to cluster binary templates of aeroplane parts in a complex hierarchical network that covers over 100,000 items, grouped into thousands of self-organised clusters. Claimed savings in manufacturing costs are in millions of dollars per annum." At Lincoln Lab, a team led by Waxman developed an image mining system which incorporates several models of vision and recognition developed in the Boston University Department of Cognitive and Neural Systems (BU/CNS). Over the years a dozen CNS graduates (Aguilar, Baloch, Baxter, Bomberger, Cunningham, Fay, Gove, Ivey, Mehanian, Ross, Rubin, Streilein) have contributed to this effort, which is now located at Alphatech, Inc.

Customers for BU/CNS neural network technologies have attributed their selection of ART over alternative systems to the model's defining design principles. In listing the advantages of its THOT® technology, for example, American Heuristics Corporation (AHC) cites several characteristic computational capabilities of this family of neural models, including fast on-line (one-pass) learning, "vigilant" detection of novel patterns, retention of rare patterns, improvement with experience, "weights [which] are understandable in real world terms," and scalability (www.heuristics.com).

Design principles derived from scientific analyses and design constraints imposed by targeted applications have jointly guided the development of many variants of the basic networks, including fuzzy ARTMAP [9], ART-EMAP [10], ARTMAP-IC [11], Gaussian ARTMAP [12], and distributed ARTMAP [3,13]. Comparative analysis of these systems has led to the identification of a default ARTMAP network, which features simplicity of design and robust performance in many application domains [4,14]. Selection of one particular ARTMAP algorithm is intended to facilitate ongoing technology transfer.

The default ARTMAP algorithm outlines a procedure for labeling an arbitrary number of output classes in a supervised learning problem. A critical aspect of this algorithm is the distributed nature of its internal code representation, which produces continuous-valued test set predictions distributed across output classes. The character of their code representations, distributed vs. winner-take-all, is, in fact, a primary factor differentiating various ARTMAP networks. The original models [9,15] employ winner-take-all coding during training and testing, as do many subsequent variations and the majority of ART systems that have been transferred to technology. ARTMAP variants with winner-take-all (WTA) coding and discrete target class predictions have, however, shown consistent deficits in labeling accuracy and post-processing adjustment capabilities. The talk will describe a recent application that relies on distributed code representations to exploit the ARTMAP capacity for one-to-many learning, which has enabled the development of self-organizing expert systems for multi-level object grouping, information fusion, and discovery of hierarchical knowledge structures. A pilot study has demonstrated the network's ability to infer multi-level fused relationships among groups of objects in an image, without any supervised labeling of these relationships, thereby pointing to new methodologies for self-organizing knowledge discovery.

Abstract: Research has been underway to examine the benefits of hybrid systems for emergent reasoning in real-time. Within this hybrid approach, there is a utilization of two types of behavioral models: explicit and intrinsic. Explicit models are identified by those actions of a robot that are predetermined by the programming and analysis that is associated with its construction. Therefore, explicit models are not learned and are used in our approach to represent responses to hard constraints (such as hitting an obstacle). Intrinsic models, on the other hand, are based upon a set of rules for activities that are highly adaptable. Such models give a robot the ability to learn how to interact with the outside world without any outside environmental sense initially built into it at all.

In one case study we performed on collision avoidance, a robot utilizes a neural network to control the direction of its motion in response to infrared sensor data. The robot learns from failure, i.e. whenever the robot collides with an obstacle, it modifies its behavior using explicit, general-purpose models (to handle hard constraints) to attempt to avoid the obstacle. Successes are trained into the neural network over time. Eventually, a coarse-grained representation of the robot's experiences emerges (in the form of k-means clusters), and these representations are paired to developing neural network strategies.

Abstract: The FOL system was designed to test some ideas about how to build thinking individuals by mechanizing the activity of reasoning and providing data structures powerful enough to represent both general information and information about reasoning itself. In this talk we will spell out the challenges implicit in the goal of FOL in some detail. Then we will describe and illustrate the basic FOL concepts: contexts, partiality, restartable computations, systems, inference rules as operations on contexts, and behaviors. Finally we will discuss how these concepts have been used to address the building of mind. (joint work with Richard Weyhrauch, IBUKI)

Abstract: How do we (humans or robots) learn about space? The Spatial Semantic Hierarchy [Kuipers, 2000] is a model of knowledge of large-scale space consisting of representations at four levels, each with its own ontology: control, causal, topological, and metrical. The SSH explains how the cognitive map can be learned from experience guided by low-level reactive control laws. How do we learn the control laws and the perceptual features they depend on? Pierce & Kuipers [1997] demonstrate an answer in terms of a radical thought-experiment. A robot with completely uninterpreted sensors and effectors moves randomly in its environment, allowing the robot to learn the structure of its sensors, generate candidate higher-level features, and identify a set of useful relationships between sensory features and motor actions. From these, homing and path-following control laws can be built: enough to support the control level of the SSH. How do we learn to recognize where we are, from the current sensory image? Sensors suffer from both perceptual aliasing (different places look the same) and image variability (the same place looks different). Kuipers & Beeson [2002] demonstrate a hybrid approach called bootstrap learning that combines unsupervised learning, deductive inference (SSH map-building), and supervised learning to reach very high accuracy place recognition in real environments using physical sensors, with very weak assumptions about the nature of the sensors.