The vacancy is focused on
calibration in deep learning. Deep Neural Networks (DNNs) have demonstrated significant predictive capabilities across various domains including computer vision, speech recognition, and natural language processing. Existing sophisticated neural network architectures are frequently integrated into practical applications. However, recent research has highlighted a crucial issue: despite their high accuracy achieved through training, deep neural networks often have overconfident and underconfident predictions. Deploying uncalibrated models in real-world systems poses substantial risks, particularly in safety-critical contexts like monitoring critical infrastructure systems. Calibrating deep learning models is essential to mitigate this risk, ensuring that the model's posterior distribution accurately represents uncertainty without being excessively overconfident.
The PhD candidate is supposed to carry out research on calibration in deep learning at the Pervasive Systems Research group, Department of Computer Science, Faculty of Electrical Engineering, Mathematics and Computer Science (EEMCS) at the University of Twente in the Netherlands. The candidate is expected to closely collaborate with the project partners such as Vrije Universiteit, Saxion, TNO-ESI and industrial partners ASML, Canon Production Printers, ITEC, Philips, and ThermoFisher Scientific.
The main research objectives are:
- Conduct research in calibrating AI, including but not limited to designing and implementing calibrated deep neural models, conducting experiments, analyzing data, and interpreting results.
- Collaborate with the team to explain the root causes of model miscalibration, optimize diagnostic processes for monitoring for Cyber-Physical System diagnostics.
- Write technical reports and research papers for publication in top-tier journals and conferences (Percom, Ubicomp, IPSN, SenSys, IJCAI, AAAI, NIPS, ICML).