Safety and Robustness for Deep Learning with Provable Guarantees - Marta Kwiatkowska - Oxford

Explore a comprehensive lecture on safety and robustness in deep learning with provable guarantees, delivered by Marta Kwiatkowska from Oxford at the Alan Turing Institute. Delve into the challenges of developing automated certification techniques for learnt software components in safety-critical applications like self-driving cars and medical diagnosis. Examine the role of Bayesian learning and causality in ensuring adversarial robustness and safety of decisions. Gain insights into emerging directions in trustworthy artificial intelligence, including machine learning accountability, fairness, privacy, and safety. Cover topics such as big data, resilience testing, adversarial perturbations, software verification, deep feedforward neural networks, robustness, certification guarantees, interventional robustness, probabilistic verification, and regression safety. Engage with the latest research and discussions on the intersection of mathematics and deep learning, addressing the need for rigorous software development methodologies in increasingly complex computing systems.

Intro
Big data
Examples
Safety
Resilience testing
Fatal crashes
An adversarial perturbation
Software verification
Machine learning
Deep feedforward neural networks
Neural networks and classifiers
Training and testing
Robustness
Safety of classification decisions
First approach
Lipsheets
Search for adversarial examples
Search for better adversarial examples
MSR for videos
Text classification
Certification guarantees
Summary
Questioning
interventional robustness
probabilistic verification
pointwise robustness
regression safety
High profile failures
We are scratching at the surface
Conclusion
Questions