Summary: This work improves AI model prediction accuracy and reliability for limited training data by developing a data-efficient Bayesian method for uncertainty quantification.

Accomplishment: Standard AI models are deterministic without uncertainty quantification (UQ), resulting in poor prediction performance and AI safety issues. Traditional UQ methods are computationally expensive and have difficulties in assessing uncertainties of AI models whose dimensionality is extremely high. In this effort, we develop a data-efficient Bayesian inference method to advance AI model optimization and uncertainty quantification. The developed Bayesian method leverages the merits of two widely used UQ methods, the Markov chain Monte Carlo sampling which is general purpose and assumption-free and the variational inference approach which converges fast to the target solution. The developed Bayesian method considers prior information which provides physical constraints leading to fast model training, it avoids overfitting which improves model prediction accuracy and robustness, and it has detailed probabilistic interpretation which supports decision making.

Significance and Impact: Scientific applications are always short of good labeled data which impairs the data-hungry AI model prediction performance. This work improves AI model prediction accuracy and reliability with uncertainty quantification for limited training data. Hence, with the paucity of labeled data, we can accurately characterize the objects and events of interest, accurately infer critical physical variables that are difficult to monitor or ahead in time, advance understanding of different physical processes and discovery of cause-effect relationships in nature sciences, and facilitate experimental design and control under uncertainty in material sciences and manufacturing.

Reference: https://ieeexplore.ieee.org/document/8955573

Funding: Research of this work is supported by LDRD AI initiative and software development is supported by DOE SciDAC project.