Neural Networks for Interactive Image and Signal Exploitation.
Abstract
Efficient reliable neural network design methods based on sound principles were developed. The main effort was concentrated on exploring the nature of error surfaces attempting to answer the question of how many minima actually exists in any given design. To this end the following experimental approach was taken: Define a scaleable model problem with a known solution, Perform comprehensive, careful training experiments, Refine training solutions to highly accurate minima, Post process and count local minima, and Derive quantitative and qualitative measures that describe error surfaces. The Principal results show: How the number of minima varies as a function of the ration of training to network size, the probability that training from a random start converges to a minimum, and How the number of minima increase with network complexity.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 29, 1995
- Accession Number
- ADA299479
Entities
People
- John Pearson
Organizations
- Sarnoff Corporation