In our CPEC Colloquium series, we invite distinguished speakers to share their research on perspicuity and related topics. We’re happy to announce the following talk:
Friday, June 18, 2:15pm: Xiaowei Huang (University of Liverpool, UK)
Safety and Reliability of Neural Networks
This talk will address the safety and reliability risks of applying neural networks to safety critical systems. It will review some of our existing research on the verification and validation of neural networks, which are mainly on the point-wise robustness property. Point-wise robustness evaluates the invariant of neural network against perturbations to an input. Based on this, it will discuss how to utilise point-wise robustness for the reliability assessment. For reliability assessment, it needs to evaluate the generalisation ability of neural networks and understands how to combine robustness and generalisation. Different from point-wise robustness, generalisation considers data samples which have not been seen during the training. It will also discuss how this reliability assessment can be applied to underwater robots.