Synchronization-Inspired Interpretable Neural Networks.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Neuroscience

Background:

• Neuronal synchronization is crucial for information processing in the human brain.
• Functional modules in the brain, like visual and motor cortices, emerge from synchronized neuronal activity.
• Existing neural networks often lack interpretability, hindering understanding of their decision-making processes.

Purpose of the Study:

• To propose an interpretable neural network model inspired by biological brain synchronization.
• To develop a mechanism that constrains neurons to capture single semantic patterns and synchronizes similar neurons.
• To enhance the interpretability of neural network representations and functional modules.

Main Methods:

• Developed a neural network incorporating a synchronization mechanism.
• Regularized neuron activation maps to focus on specific pattern positions.
• Implemented adaptive synchronization of locally interacting neurons during training.
• Introduced novel evaluation metrics for comprehensive neuron interpretability analysis.

Main Results:

• The proposed method significantly improved neuron interpretability compared to state-of-the-art algorithms.
• Synchronized functional modules demonstrated consistency across different datasets.
• Individual modules exhibited high semantic specificity.
• Qualitative and quantitative experiments validated the effectiveness of the approach.

Conclusions:

• The synchronization mechanism effectively promotes the formation of interpretable functional modules in neural networks.
• The model achieves a balance between preserving distributed representations and enabling interpretable modules.
• This approach offers a promising direction for developing more transparent and understandable artificial intelligence systems.