Learning activation rules rather than connection weights

Area of Science:

• Artificial Intelligence
• Computational Neuroscience

Background:

• Neural networks often benefit from local representations for associative recall.
• A priori knowledge can guide the selection of connection weights.
• Standard learning algorithms fail when connection weights are fixed.

Purpose of the Study:

• To develop a supervised learning rule for training neural networks with fixed connection weights.
• To enable network training by modifying activation rules instead of connection strengths.

Main Methods:

• Derived a supervised learning rule based on gradient descent.
• Incorporated both traditional and competitive activation mechanisms.
• Implemented the learning rule and tested it on several networks.

Main Results:

• The novel learning rule effectively trains neural networks with fixed connection weights.
• Competitive activation mechanisms proved efficient for network training.
• Empirical results demonstrate the learning rule's effectiveness.

Conclusions:

• The proposed learning rule offers a viable solution for training neural networks with predetermined connection weights.
• Adapting activation rules is a powerful alternative to modifying connection strengths for network training.
• This approach enhances the flexibility and applicability of neural network training.