About
PhD Candidate at The University of Queensland, working on neural network verification and, with collaborators, on model and data control and on privacy.
Position
I am a PhD Candidate at The University of Queensland, supervised by A/Prof. Guangdong Bai. My work is on making claims about neural networks that can be checked rather than measured: what can be proved about a network over a bounded input set, and what has to be assumed to prove it.
My main research area is neural network verification, and the doctoral work is about the convex approximations that sound bounds on model behaviour are built from. WraLU and WraAct study efficient approximations of activation-function hulls.
Alongside that I collaborate on model and data control, and on privacy in learning and retrieval. The research page brings those themes together, and every project page credits all coauthors and links to the original paper and implementation.
Research lines
- Formal Verification & Robustness
- What can be proved about a neural network over a specified set of inputs? (3 works)
- Model & Data Control
- Once a model or its data has been released, how can its capability and their utility still be constrained? (4 works)
- Privacy in Learning & Retrieval
- What information escapes through gradients and embeddings, and how much of it can be taken back? (3 works)
Where things are
- PhD
- PhD Candidate, The University of Queensland
- Supervisor
- A/Prof. Guangdong Bai
How this got here
I also keep a chronological account of the journey -- what the hard parts were, what compounded, and which research directions came out of it.