Entropy-based generation of supervised neural networks for classification of structured patterns

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Generalized recursive neural networks offer a novel approach to structure classification.
• Existing methods for constructing these networks can be complex and may not yield optimal performance.

Purpose of the Study:

• To propose an automated, entropy-based method for constructing generalized recursive neural networks.
• To enhance the classification of acyclic structured patterns.

Main Methods:

• An entropy-based approach is utilized to determine neural network architecture (hidden layers, neurons per layer).
• Link weights within the neural network are automatically calculated.
• The method focuses on the classification of acyclic structured patterns.

Main Results:

• The proposed method automatically determines network architecture and link weights.
• Networks built using this approach demonstrate superior performance compared to other methods.
• Improvements are noted in network size, learning speed, and recognition accuracy.

Conclusions:

• The entropy-based construction method provides an effective way to build generalized recursive neural networks.
• This automated approach leads to improved efficiency and accuracy in classifying structured patterns.