Exploring how deep learning decodes anomalous diffusion via Grad-CAM.

Area of Science:

• Physics
• Computer Science
• Data Science

Background:

• Deep learning excels at classifying anomalous diffusion from trajectory data.
• The internal decision-making processes of these deep learning models remain largely unexplained.
• Understanding these mechanisms is crucial for improving model reliability and interpretability.

Purpose of the Study:

• To investigate how deep learning models, specifically ResNets, recognize features of anomalous diffusion mechanisms.
• To utilize explainable AI techniques, like Gradient-weighted Class Activation Map (Grad-CAM), to visualize and understand the model's feature recognition process.
• To determine if insights from explainable AI can enhance classifier robustness against noise.

Main Methods:

• Implemented deep learning models (ResNets) for anomalous diffusion classification.
• Applied Gradient-weighted Class Activation Map (Grad-CAM) to analyze the ResNets' decision-making process.
• Examined the model's focus on specific trajectory segments and spatiotemporal scales.

Main Results:

• Grad-CAM successfully identified trajectory portions critical for classifying anomalous diffusion mechanisms.
• The identified crucial features can be used to improve classifier robustness against measurement noise.
• Deep learning models were found to capture distinct statistical characteristics of different diffusion mechanisms across various spatiotemporal scales.

Conclusions:

• Explainable AI (Grad-CAM) provides insights into deep learning's recognition of anomalous diffusion features.
• Understanding feature importance enhances model robustness and interpretability.
• Deep learning effectively distills multi-scale statistical information from trajectory data.