Theory and Applications of Neural Networks
Abstract
One of the main ideas underlying the interest in neural computing is that it may be possible to develop new computational paradigms that will make important aspects of programing both simple and more robust. The means for doing so usually involves usually involves setting up some universal difference or differential equation whose trajectories define rules for solving problems in curve fitting, interpolation, etc. The work has addressed the use of analog computation methods for optimization as well as sorting, quantizing, etc. Using a simple, but powerful, mathematical model they have shown, how basic subsystems can provide the building blocks that are capable of accounting for the operations that they see being performed by biological and digital computers. More specifically, they have shown that a certain class of gradient flows on the n dimensional orthogonal group generates effective means for solving a variety of combinatorial and linear algebra problems of the type that shows up in the neural network literature. A key idea here is that of ail adaptive subspace filter - a general model for nonlinear filtering of the type seen in various cognitive applications. This model not only allows one to study global convergence in a precise way, but it allows one to make analytical predictions about the speed of convergence which then can be compared with the performance of natural systems. They have shown that some of the earlier analog models for sorting can be interpreted as conditions density propagators.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 28, 1992
- Accession Number
- ADA262890
Entities
People
- Roger Brockett
Organizations
- Harvard University