Improving the Interpretability of fMRI Decoding using Deep Neural Networks and Adversarial Robustness.

Area of Science:

• Neuroscience
• Machine Learning
• Data Science

Background:

• Deep neural networks (DNNs) are powerful tools for analyzing complex, high-dimensional data but are often considered "black boxes" due to their lack of interpretability.
• Understanding which input features drive DNN predictions is crucial for applications in cognitive neuroscience and neuroinformatics.
• Saliency maps are commonly used to visualize input feature importance, but existing methods can be unreliable, sensitive to noise, and difficult to evaluate.

Purpose of the Study:

• To review gradient-based saliency map methods for DNNs.
• To introduce a novel adversarial training method to enhance DNN robustness to input noise and improve interpretability.
• To propose and validate quantitative evaluation procedures for saliency map interpretability in neuroimaging data.

Main Methods:

• A review of existing gradient-based saliency map techniques.
• Development of an adversarial training approach for noise-robust DNNs.
• Introduction of two quantitative evaluation procedures for saliency maps, tested on synthetic data and functional magnetic resonance imaging (fMRI) data from the Human Connectome Project (HCP).

Main Results:

• Saliency maps generated by different methods exhibit significant variability in interpretability.
• Despite comparable decoding performance, DNN-derived saliency maps showed higher interpretability than those from linear models.
• The proposed adversarial training method resulted in saliency maps that outperformed alternative methods in interpretability.

Conclusions:

• The interpretability of saliency maps varies widely across different methods and model types.
• Adversarial training offers a promising approach to enhance the interpretability of DNNs in neuroimaging.
• The developed evaluation procedures provide a robust framework for assessing saliency map quality in decoding tasks.