Postdoc positions in Safety Assurance for Machine Learning @ NTU, Singapore

Position type: Postdoc (Research Fellow) for 3 years.

Salary range: Between S$4500 and S$6000 monthly depending on background and experience.

Expected qualifications: PhD in Computer Science, Computer Engineering or related fields.

Number of positions: 2

Research topics:

1. High assurance techniques for efficient runtime detection of out-of-distribution images in vision based applications for CPS (training space characterization).

2. Probabilistic formal verification techniques for safety assurance of learning enabled components in CPS (e.g., data-driven techniques based on scenario optimization).

Project description:

Modern cyber-physical systems (CPS) such as autonomous vehicles and smart grids are expected to perform complex control functions under highly dynamic operating environments. Increasingly, to cope with the complexity and scale, these functions are being designed autonomously, often with the assistance of Machine Learning (ML) based techniques. A key challenge in the design of such safety-critical systems is that of achieving high assurance on system safety. One aspect of this challenge is that the ML models are complex and process extremely high-dimensional inputs such as images, thus rendering the resulting black-box controllers inaccessible to traditional safety assessment techniques. Another aspect of this challenge is that the correctness of such learning algorithms is critically dependent on the training data sets. In the case of CPS, these sets are often non-representative and missing rare, but important, safety-related hazard data.

Addressing the above two open problems, in this project we aim to develop a software architecture for achieving safety assurance in such systems. This architecture is inspired by the Simplex model of computation in which a complex and high-performance controller is coupled with a simple and conservative controller to ensure system safety. An efficient run-time decision module will then switch between the two controllers while ensuring safety at all times. To develop a verifiable safety controller, we aim to focus on scalable techniques with probabilistic correctness (probably approximately correct - PAC) guarantees for example using scenario optimization. For the run-time decision module, we aim to focus on statistically robust and computationally efficient techniques for out-of-distribution detection to quickly identify and deal with images that are outside the training space of learning enabled components.

Research environment:

In this position, you are expected to collaborate with and manage a team of PhD students and research assistants; the latter will support implementation activities in the project including a simulation platform (e.g., Carla) and an embedded testbed (e.g., Duckietown). Opportunities for teaching can also be facilitated depending on interest.

This position is part of a well-established and globally renowned research team in CPS (https://cps-research-group.github.io/), in the School of Computer Science and Engineering (http://scse.ntu.edu.sg/Pages/Home.aspx) at Nanyang Technological University (https://www.ntu.edu.sg/Pages/home.aspx).

If interested, please send an email along with a detailed CV to Arvind Easwaran (arvinde@ntu.edu.sg).