George Berg Ph.D., Northwestern University
Associate Professor
(Personal Page)
Computer Science Department
University at Albany
Albany, NY 12222
(518) 442-4267 
(518) 442-5638 (FAX)
Personal Statement of Research
The theme of my research is connectionist or artificial neural network models of problems involving complex, structured information. The methodology is to devise network architectures which can be trained on examples from the problem domain in order to produce a network which solves the problem in general. Specifically, my interests are in two areas, natural language processing and molecular biology.

In the natural language processing research, I have devised a connectionist parsing model, XERIC. XERIC combines Simple Recurrent Networks, RAAM-style reduced encodings, and a novel training regimen using virtual copies of the network. The resulting network is able to parse sentences of English of both active and passive form, as well as those with recursively embedded phrases and embedded sentences. These may be of arbitrary length and complexity, limited only by the ability of the reduced representations to encode information.

In molecular biology, the specific application is the protein folding problem. Proteins are made up of sequences of amino acids, of which there are twenty different types. A protein folds up into a characteristic three-dimensional shape. This shape helps determine the protein's function. In theory this shape can be predicting knowing only the protein's sequence. In practice, this is an open problem. It is important because experimentally determining a protein's shape is a costly and time-consuming laboratory process. In collaboration with Professor Jacquelyn Fetrow of the Department of Biological Sciences, we are using recurrent network architectures in order to build networks which predict proteins' shapes given only their amino acid sequences. 

Follow this link for more details.

Selected Publications
 "Design of an Auto-Associative Neural Network with Hidden Layer Activations that were used to Reclassify Local Protein Structures", with X. Zhang and J. S. Fetrow, in Techniques in Protein Chemistry V, J. W. Crabb, (Ed.), pp. 397-404, 1994.
"Automatic Derivation of Substructures Yields Novel Structural Building Blocks in Globular Proteins", with X. Zhang, J. S. Fetrow, D. L. Waltz and W. A. Rennie, Proceedings of the First International Conference on Intelligent Systems for Molecular Biology, pp. 438-446, July 1993.
"A Connectionist Parser with Recursive Sentence Structure and Lexical Disambiguation", Proceedings of the Tenth National Conference on Artificial Intelligence (AAAI-92), pp. 32-37, July 1992.
"The Case for Non-connectionist Associative Retrieval in Case-Based Reasoning Systems", with P. Bonissone and L. Rau, in Advances in Connectionist and Neural Computation Theory; Volume 2: Connectionist Approaches to Analogy, Metaphor and Case-Based Reasoning, K. Holyoke and J. Barnden (Eds.). In press.
"Representational Adequacy and the Case for a Hybrid Connectionist/Marker-Passing Model", in Connectionist Approaches to Language Processing, R. Reilly and N. Sharkey (Eds.). Lawrence Erlbaum, pp. 253-272, 1992.
"A Connectionist Model of Motion and Government in Chomsky's Government-Binding Theory", with J. E. Rager, Connection Science, special issue on "Connectionist Natural Language Processing", Vol. 2, pp. 35-52, 1990. Reprinted in Connectionist Natural Language Processing: Readings from Connection Science, N. Sharkey (Ed.). pp. 28-45, Kluwer, 1992.