Bayesian deep learning for error estimation in the analysis of anomalous diffusion.

Area of Science:

• Physics
• Biophysics
• Ecology

Background:

• Single-particle tracking generates vast time-series data of diffusive motion across diverse systems.
• Understanding the physical mechanisms behind this motion is crucial for system comprehension.

Purpose of the Study:

• To enhance machine-learning models for anomalous diffusion data analysis.
• To incorporate uncertainty quantification into predictions for improved reliability.

Main Methods:

• Utilized Bayesian-Deep-Learning, specifically Stochastic-Weight-Averaging-Gaussian (SWAG).
• Trained models for diffusion model classification and anomalous diffusion exponent regression.
• Evaluated model performance on accuracy and calibration of error estimates.

Main Results:

• Achieved well-calibrated uncertainty estimates alongside high prediction accuracies.
• Demonstrated the models' ability to quantify prediction reliability.
• Related output uncertainty to underlying diffusion model properties.

Conclusions:

• The enhanced machine-learning approach provides reliable predictions and uncertainty quantification for anomalous diffusion.
• This method offers insights into both the physical systems and the machine learning model's behavior.
• The technique is applicable to single-particle trajectories in various scientific domains.