C-Mantec: a novel constructive neural network algorithm incorporating competition between neurons.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Existing constructive neural network algorithms face challenges in stability and adaptability.
• Neural network architecture growth and neuron learning rules require careful consideration for effective knowledge acquisition.

Purpose of the Study:

• Introduce C-Mantec, a novel constructive neural network algorithm.
• Evaluate C-Mantec's performance in terms of architecture size and generalization ability.
• Address the issue of overfitting in neural networks.

Main Methods:

• C-Mantec combines neuron competition with a modified perceptron learning rule (thermal perceptron rule).
• The algorithm was tested on Boolean functions for logic circuit design and real-world benchmark datasets.
• A built-in method for avoiding overfitting was developed and applied within an active learning paradigm.

Main Results:

• C-Mantec generated highly compact neural network architectures.
• The algorithm demonstrated state-of-the-art generalization capabilities across benchmark problems.
• The built-in method effectively filtered noisy examples and mitigated overfitting.

Conclusions:

• C-Mantec offers a stable and adaptive approach to neural network construction.
• The algorithm achieves superior generalization performance compared to standard classification methods.
• C-Mantec provides an effective solution for building efficient and accurate neural network models.