Connectionist Models for Intelligent Computation
Abstract
(1) Research Objectives: -- To study the underlying principles, architectures and applications of artificial neural networks for intelligent computations. (2) Approach: -- We use both numerical simulation and theoretical analysis to investigate various alternatives in connection schemes, organization principles and architectures of artificial neural networks. (3) Progress for period 9/1/88-8/31/89: -- In the past year, our research on neural network models for intelligent computing under the sponsorship of AFOSR continues to make important progress. In particular, we have constructed the Parallel Sequential Induction Network, a powerful network that self-organizes into an optimal structure to perform classification tasks. In neural network research, much attention has been paid to improving the efficiency of learning connection weights for a network with fixed topology. However, little progress has been made toward uncovering optimal designing principles to reshape the connection topology of a network adaptively to maximize the performance of a specific task. Recent studies indicate that multi-layered feedforward networks of sufficient complexity, in general, need only two hidden layers to imitate any decision hypersurface in the pattern space. (kr)
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 26, 1989
- Accession Number
- ADA228949
Entities
People
- H. H. Chen
- Y. C. Lee
Organizations
- University of Maryland