Related Experiment Videos
Engineering multiversion neural-net systems
1Department of Computer Science, University of Exeter, England.
Neural Computation
|May 15, 1996
Summary
This study enhances neural network reliability by using diverse, multiversion systems with decision strategies. This approach minimizes errors from individual component failures, improving overall system performance and reliability.
Area of Science:
- Computer Science
- Artificial Intelligence
- Software Engineering
Background:
- Neural network implementations often contain errors.
- Ensuring the reliability of complex neural network systems is a significant challenge.
Purpose of the Study:
- To develop and evaluate methods for constructing reliable neural network systems.
- To leverage multiversion system principles to mitigate errors in individual neural network components.
Main Methods:
- Implementing and testing diverse, multiversion neural network systems.
- Utilizing a "decision strategy" (e.g., majority vote) to combine outputs from multiple versions.
- Evaluating techniques for selecting optimal subsets of system components from an overproduced set.
Main Results:
- Demonstrated that methodological diversity improves system performance.
- Evaluated three component selection techniques for optimal system engineering.
- Assessed the reliability of complex multiversion designs compared to individual component reliability.
Conclusions:
- Multiversion systems, exploiting methodological diversity, can significantly enhance neural network reliability.
- The proposed approach offers a viable strategy for building robust and dependable neural network applications.