Gail Carpenter - US grants

CARPENTER The investigator intends to develop and analyze models of neural systems networks. These models will be designed to carry out a self-organizing pattern recognition and category learning. As formal architectures, the models will be translated into machine designs. As neural networks, the systems will be based on psychological and physiological data and will be used to analyze the brain and behavior in the process of pattern recognition and category learning. From a mathematical point of view, these models are based on nonlinear, singular, high dimensional systems of ordinary differential equations. Network design is constrained by the conflicting requirements of stability and adaptability, and by the complexity. Given an arbitrary complicated sequence of inputs, the system must be able to extract stable categories, without a teacher. However, the system must also remain plastic, open to new learning. New network designs will generalize systems which have been shown to stably encode arbitrary lists of binary input patterns in such a way that, after an initial phase, each input directly activates its category representation. The search mechanism is automatically disengaged once the learning is complete. The more general system will encode graded input patterns, and be embedded in network hierarchies adapted to the particular constraints of natural and artificial vision and speech. Research described in this project is highly cross-disciplinary, as it involves the neural networks science, communications, some aspects of robotics and computer vision, neural cognitive processes, and mathematics.

Dr. Carpenter and two graduate students will further develop and study her biologically motivated ART3 model of adaptive pattern recognition. The funded work will focus on computational analysis of these self-organizing architectures, development of network architectures for new parallel computer designs, and construction of hierarchies of neural-network modules (including her earlier ART and ART2 "adaptive resonance theory" networks).

9401659 Carpenter The proposed research will design, develop, analyze, and apply neural network architectures for stable recognition coding and prediction. One set of projects will focus on new unsupervised ART (Adaptive Resonance Theory) and supervised ARTMAP systems to support distributed recognition codes that are stable under training regimes with either fast or slow learning. Typically, the pattern of activity of a recognition code is distributed when there is little reason to select one category over another, as might occur early in a learning process. This pattern becomes more compressed as categories are more sharply defined through experience. Generalization and storage capacity of the ART neural networks will be greatly increased by the distributed coding capability. Related work will focus on the computational requirements of distributed search and on spatio-temporal recognition problems.

The Center of Excellence for Learning in Education, Science, and Technology (CELEST) brings together leading scientists, educators, and technologists from Boston University, Brandeis University, Massachusetts Institute of Technology, and the University of Pennsylvania to study real-time autonomous learning systems by integrating experimental and computational brain science, biologically inspired technology, and classroom innovation. Contributing scientists are drawn from four Boston University Departments and the Center for Adaptive Systems, the Center for Memory and Brain, the Science and Mathematics Education Center, the Hearing Research Center, and the Center for Polymer Studies; the Brandeis University Department of Psychology and the Volen Center for Complex Systems; the MIT Department of Brain and Cognitive Sciences, the Picower Center for Learning and Memory, and the Harvard/MIT Speech and Hearing Bioscience and Technology Program; and the University of Pennsylvania Department of Psychology.
Intellectual Merit and Creative Concepts: CELEST brings together educators, scientists, and technologists to carry out four types of mutually reinforcing and integrated activities: (1) quantitative behavioral and brain modeling of both normal and abnormal learning processes during perception, cognition, emotion, and action; (2) interdisciplinary cognitive and neuroscience experiments to probe these processes and to test model predictions; (3) development of algorithms, based on biological learning models, for incremental fast learning about complex and rapidly changing environments in large-scale engineering and technological applications that are important in many areas of society; and (4) integration of research and education through contributions to educational technology, curriculum development, and early career recruitment of underrepresented communities into scientific practice. These goals are achieved through interactions among eight main Thrusts in:
Learning in visual perception and recognition: laminar cortical dynamics of adaptive behavior;
Learning in audition, speech, and language; learning in cognitive-emotional interactions and planned sequential behaviors;
Learning and episodic memory: encoding and retrieval;
Learning in concept formation and rule discovery;
Learning in attentive recognition and neuromorphic technology;
Educational technology, curriculum development, and outreach; and
Diversity outreach.
Broader Impact: CELEST will foster interdisciplinary collaborations and training across all its units: frequent seminars, workshops, colloquia, conferences, and publications; integration of research and education by translating basic science results into interdisciplinary curriculum development; and web-based and hands-on training for teachers and students, including classroom activities with a national and international impact. CELEST will hereby provide world-class expertise towards advancing key SLC program goals; namely, the psychological and pedagogical aspects of learning, the biological basis of learning, machine learning, learning technologies, and mathematical analyses and modeling of them all.