Accession Number:

ADA237856

Title:

Analog Computation in Neutral Systems: Architectures and Complexity

Descriptive Note:

Final technical rept. 1 Aug 1988-31 Jul 1990,

Corporate Author:

PRINCETON UNIV NJ DEPT OF ELECTRICAL ENGINEERING AND COMPUTER SCIENCE

Personal Author(s):

Report Date:

1991-05-17

Pagination or Media Count:

12.0

Abstract:

First, we studied the representation problem for the class of single- hidden-layer feedforward networks, which is fundamental for understanding limitations of learning algorithms, and which also contributed to understanding the behavior of learning algorithms in applications involving low-complexity networks. The second kind of problem studied concerns dynamics behavior in neural networks containing feedback trellis-structured networks in one particular applications. Our work focused on studying stability issues and exploring the implications of computational complexity theory. Third, the PAC learning paradigm probably Almost Correct was analyzed with the goal of characterizing the effects of statistically dependent sequences of training examples on learning performance. The goal of all these efforts was to discover and explore insights about fundamental limitations on the computational capabilities of analog neural systems and, where possible, of more general classes of physical systems as well.

Subject Categories:

  • Operations Research

Distribution Statement:

APPROVED FOR PUBLIC RELEASE