Transformations of sigma-pi nets: obtaining reflected functions by reflecting weight matrices

Area of Science:

• Artificial Intelligence
• Computational Neuroscience

Background:

• Trained artificial neural networks typically have fixed input-output mappings.
• Adapting networks to new tasks often requires extensive retraining.

Purpose of the Study:

• To present a methodology for transforming the functionality of post-trained artificial neural networks.
• To enable networks to associate different input-output mappings without retraining.

Main Methods:

• The methodology manipulates the weight matrices of sigma-pi neural networks.
• Transformations are performed on the weight matrix, not through retraining the network.
• Specific transformations include reflections across horizontal and vertical axes and scaling.

Main Results:

• The transformed weight matrices successfully altered the network's output function.
• A reflection transformation on the weight matrix resulted in a vertical axis reflection of the function.
• The network performed related mapping tasks after learning an initial task.

Conclusions:

• Weight matrix transformation offers a method to adapt trained neural networks to new tasks efficiently.
• This approach enhances the flexibility and applicability of artificial neural networks.
• Further research can explore deriving a set of standard transformations for sigma-pi networks.