Interpretable Neuron Structuring with Graph Spectral Regularization.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Neural networks are powerful function approximators but often function as black boxes.
• Interpreting the internal workings and features of hidden layers remains a significant challenge.
• Understanding neuron activation patterns is crucial for biological plausibility and model interpretability.

Purpose of the Study:

• To develop a method for enhancing the interpretability of neural network hidden layers.
• To introduce Graph Spectral Regularization (GSR) as a technique to structure neuron activations.
• To maintain high performance on primary tasks while improving model transparency.

Main Methods:

• Implemented Graph Spectral Regularization using a graph Laplacian penalty.
• Structured hidden layer activations based on spatial organization principles from biological networks.
• Explored two graph construction approaches: predetermined graphs and data-driven feature-space graphs derived from co-activations.
• Applied the regularization to both biological and image datasets.

Main Results:

• Demonstrated that Graph Spectral Regularization enhances the interpretability of hidden layer features.
• Showcased the ability of the method to reveal cluster indications within the data.
• Validated the effectiveness of the approach through visualization techniques on diverse datasets.
• Confirmed that interpretability gains were achieved without significant degradation of primary task performance.

Conclusions:

• Graph Spectral Regularization offers a viable solution for the interpretability challenge in neural networks.
• The method provides valuable insights into data structure and neuron function.
• The technique holds promise for applications in fields requiring transparent AI models, such as bioinformatics and computer vision.