reynolds@park.bu.edu (John Reynolds) (05/23/91)
Hi, Surya. Recently, Stephen Grossberg, Gail Carpenter and I have introduced a supervised Adaptive Resonance Theory (ART) network, called ARTMAP, which can use knowledge about class membership to cluster patterns meaningfully. On a trial by trial basis it develops a highly efficient tessellation of the input space which maps each input cluster onto an appropriate output cluster. In the case of known class membership, each distinct output vector could merely be a label of class membership. In the more general case, the output vectors reflect central tendencies or critical feature patterns of output vector clusters. THE BASIC ALGORITHM: Each time an input pattern is assigned to a cluster corresponding to an output cluster that does not include the output vector, the input tessellation is changed by the minimum amount necessary to correctly classify the input vector. Specifically, the system automatically identifies clusters which (1) are associated with the output cluster containing the output pattern and (2) will, after learning, code the current input. If such clusters exist, it identifies the one that will undergo the minimum change in order to claim the current input pattern. If there are no such clusters, it creates a new cluster which codes the current input and associates it with the correct output cluster. In this way it can classify arbitrarily many, arbitrarily ordered vectors into recognition categories based on predictive success. It conjointly maximizes predictive generalization and minimizes predictive error by linking predictive success to category size on a trial-by-trial basis, using only local operations. The system was presented a couple of weeks ago at the Wang Institute Conference on Neural Networks for Vision and Image Processing, and it will also appear at the upcoming IJCNN meeting (Lecture, Friday, July 12, Session 2, 9:10 - 9:30AM). It will be discussed in an upcoming issue of Neural Networks (Neural Networks, 4, in press), and it is now available as Technical Report CAS/CNS-TR-91-001. If you would like a copy of the technical report, please write to the following address: Boston University Center for Adaptive Systems and Cognitive and Neural Systems Department 111 Cummington Street, Rm. 244 Boston, MA 02215 or contact Cindy Suchta (cindy@park.bu.edu) to request a copy of the technical report. Be sure to include the technical report number. -John Reynolds -- =========================== MODERATOR ============================== Steve Stevenson {steve,fpst}@hubcap.clemson.edu Department of Computer Science, comp.parallel Clemson University, Clemson, SC 29634-1906 (803)656-5880.mabell