Redesigning deep neural networks: Bridging game theory and statistical physics.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Physics

Background:

• Traditional deep learning models face challenges in unified feature extraction and pattern classification.
• Integrating principles from game theory and statistical physics offers a novel approach to enhance neural network capabilities.

Purpose of the Study:

• To propose and validate a hybrid deep-graphical representation combining game theory and statistical physics.
• To improve feature extraction, pattern classification, model regularization, and explainability in deep learning.

Main Methods:

• Neurons are modeled as players in game theory and particles in statistical physics.
• Neural network layers are treated as sequential cooperative games.
• Shapley value and Banzhaf Power Index are used for neuron assessment, filtering, and explainability.

Main Results:

• The hybrid model demonstrated superior efficiency and accuracy compared to traditional multilayer perceptron and convolutional neural network models.
• Neurons with significant contributions form coalitions, enhancing information transmission.
• Shapley value improved model regularization and overall performance.

Conclusions:

• The proposed deep-graphical representation offers a powerful hybrid framework for advanced deep learning tasks.
• This approach enhances model performance, regularization, and explainability.
• The model shows significant potential for applications in facial age estimation and gender classification.