Tianle Zhang is currently a PhD student supervised by Dr Wenjie Ruan and Prof. Jonathan Fieldsend at the Exeter Trustworthy AI Lab, University of Exeter. Previously, he received his MSc in Statistics and BSc in Information and Computing Science from the School of Mathematics and Statistics, Central South University, Changsha, China.
Tianle Zhang's current research interest is the safety and trustworthiness of Deep Neural Networks (DNNs) with provable guarantees. Specifically, his research focuses on verification methods based on statistical insight to evaluate the reliability of DNNs under uncertainties with provable guarantees. On the other hand, his research aim is concerned with improving the robustness and safety against input perturbations for safety-critical applications.
Deep learning has attracted worldwide attention and has been applied in many scenarios, including image classification, machine reading, self-driving cars, etc. However, employing this approach in systems is not without risks, especially in safety-critical systems. For instance, a red traffic light can be falsely recognised as a green light by changing only one pixel in a self-driving system. Therefore, before applying and deploying any deep learning-based components into a safety-critical system, it is required that deep learning techniques must be accurate, robust and trustworthy. To utilise deep learning in the safety-critical industries with confidence, how to verify and improve the trustworthiness of DNN is the focus of his research.
Specifically, he is currently working on verification methods based on statistical insight to evaluate the reliability of DNNs under uncertainties, i.e., to verify the probabilistic robustness of DNNs, which means the network is deemed as probabilistically robust if the probability of failure is lower than a critical level. Moreover, he will enhance DNNs’ robustness and safety to tackle one challenge, i.e., the unfavourable trade-off between accuracy and robustness according to chance-constrained optimisation. In particular, he tries to bridge the gap between the accurate yet brittle average-case and the robust yet conservative worst-case by enforcing robustness to most rather than all perturbations, that is, probabilistic robustness.