Related Experiment Videos
An iterative method for training multilayer networks with threshold functions.
E M Corwin1, A M Logar, W B Oldham
1Dept. of Math. and Comput. Sci., South Dakota Sch. of Mines and Technol., Rapid City, SD.
IEEE Transactions on Neural Networks
|January 1, 1994
Summary
This study introduces a method for training feed-forward neural networks using threshold functions, reducing hardware complexity. The technique gradually transforms logistic functions into threshold functions, enabling efficient offline learning for hardware implementations.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Neural Networks
Background:
- Feed-forward networks are crucial in machine learning.
- Hardware implementation complexity is a challenge.
- Threshold functions offer reduced hardware complexity compared to logistic functions.
Purpose of the Study:
- To develop a mathematical foundation for training neural networks with threshold functions.
- To enable efficient hardware implementation of trained networks.
- To bridge the gap between logistic and threshold function networks.
Main Methods:
- Gradually altering logistic function gain parameters towards infinity.
- Utilizing backpropagation with gradient descent for initial training.
- Progressively increasing the steepness of logistic functions.
Main Results:
- Demonstrated that small changes in gain parameter cause small changes in network error.
- Successfully mapped trained weights to a threshold function network in simulations.
- Validated the feasibility of the gradual transformation method.
Conclusions:
- The proposed method allows training networks with logistic functions and mapping them to threshold function networks.
- This approach is suitable for offline learning scenarios where hardware implementation is a priority.
- The technique offers a practical pathway to simplified hardware neural network designs.
