Evaluation of Uncertainty Quantification (UQ) methods in Deep Learning (DL) Models for Healthcare Cyber-Physical System (CPS) Data

Master

In recent years, DL models are widely being adopted in a plethora of applications, such as smart power grids, robotics, autonomous vehicles, and so forth. However, such models are always accompanied with inherent uncertainty that needs to be appropriately handled to prevent unsafe behavior of the CPSs under test. Several different approaches have been proposed in the literature that address the topic of UQ in DL [1]. However, due to the high complexity of both CPSs but also DL algorithms (e.g., model hyperparameters), identifying the most efficient UQ approach for a given application/dataset is not a simple task. In this thesis, you will focus on evaluating a selected number of UQ methods using CPS data from the healthcare domain aiming to provide insights on the performance of these algorithms in both classification and regression based (time-series) problems.

Goal

  • Study and implementation of state-of-the-art UQ methods in DL.
  • Assessing their application in healthcare CPS data by using statistical analysis tools.

Learning outcome

  • Increase your knowledge in ML/DL concepts and UQ methods.
  • Enhance your programming skills and strengthen your background on statistical approaches for data analysis.

Qualifications

  • Background in programming languages such as R/Python.
  • Background in ML/DL algorithms.
  • Familiarity with libraries such as Keras and Tensorflow.

Supervisors

References

[1] Abdar, Moloud, et al. "A review of uncertainty quantification in deep learning: Techniques, applications and challenges." Information Fusion 76 (2021): 243-297.

Contact person