My research is supported in part by CELEST, an NSF Science of Learning Center (NSF SBE-0354378)
I first read about Adaptive Resonance Theory (ART) when looking for a nice thesis topic as an electronics engineering undergrad. At that time, Stephen Grossberg's models of vision and learning pushed my way into graduation. More than eight years later, I find myself working with Dr. Gail Carpenter (Steve's wife) to further develop ART models and explore technological applications that can take advantage of their unique mechanisms.
Our models explore applications in information fusion and knowledge discovery, bringing ideas of animal visual processing, learning, and recognition to models of image processing, learning, and pattern recognition and classification.
Self-organizing knowledge discovery
Knowledge discovery mechanisms allow a recognition and classification system to extract underlying associations among the classes it learns to identify. These mechanisms can be used to construct a hierarchical knowledge structure that represents information embedded in the processed data. This knowledge structure has the potential of taking an active role in the learning process; for example, the system could use recently acquired knowledge to improve its category learning. Designing a system that simultaneously learns to classify while extracting a knowledge structure raises questions such as how early in the learning process self-organized associations will be reliable enough to discover relationships among the classes.
Recognition and classification systems capable of discovering a knowledge structure in training data may take advantage of newly discovered rules to improve their performance. Design questions include when, how, and to what extent the continuously evolving knowledge structure should affect classification learning. Additionally, should the enhancement process modify previously learned elements, create new ones, or only bias the classification result?
Distributed internal representation of information in a recognition and classification system is essential for both self-organizing knowledge discovery and rule-based inference. In order to integrate benefits of rule-based inference into systems using ART, without sacrificing memory compression, stability, or fast online learning, distributed learning rules need to meet certain criteria. Distributed code representations should be used as early as possible during training to enable rule extraction and rule-based inference; but how is it possible to guarantee a stability-plasticity balance, in particular at initial stages during which most of the critical learning takes place and distributed predictions are not necessarily accurate?