A Complexity Theory of Neural Networks
Abstract
Significant progress has been made in laying the foundations of a complexity theory of neural networks. The fundamental complexity classes have been identified and studied. The class of problems solvable by small, shallow neural networks has been found to be the same class even if (1) probabilistic behaviour (2)Multi-valued logic, and (3)analog behaviour, are allowed (subject to certain resonable technical assumptions). Neural networks can be made provably fault-tolerant by physically separating the summation units from the thresholding units. New results have also been obtained on the complexity of approximation, communication complexity, the complexity of learning from examples and counterexamples, learning with multi-valued neurons, exponential lower bounds for restricted neural networks, and fault tolerance in distributed computation.
Document Details
- Document Type
- Technical Report
- Publication Date
- Aug 09, 1991
- Accession Number
- ADA241807
Entities
People
- Georg Schnitger
- Ian Parberry
- Piotr Berman
Organizations
- Pennsylvania State University