Vu-Anh Le

I build mathematical tools that make AI systems dependable and understandable. My work uses ideas from topology, graph theory, and statistics to provide guarantees about when models will behave stably and why they make certain predictions. I then apply these tools to real problems, including climate and natural-hazard risk.

1. Robustness & Scalability

My work on robustness and scalability keeps AI models stable when their data or connections change and ensures they continue to perform as problems grow. I study how small edits or noise influence predictions and aim to make those effects predictable and limited. Using tools such as Laplacian eigenvalues, diffusion operators, and persistent homology, I quantify sensitivity to perturbations and translate these measurements into designs that remain reliable and efficient on large, evolving datasets.

2. Interpretability & Diagnostics

I develop methods that make model behavior transparent to users and provide early warnings when predictions may be untrustworthy. By building structure-aware explanations and lightweight diagnostic tests that run during training or inference, I reveal why a model produced a given output and flag cases where the model is likely to fail, enabling safer decision-making.

3. Graph Learning

I analyze learning on networks where relationships carry as much information as features. Focusing on graph filters and graph neural networks—including settings with complex geometry such as δ-hyperbolic graphs—I establish conditions under which small, local changes can alter or preserve global predictions. These guarantees help practitioners understand and control model behavior on real-world, dynamic graphs.

4. Efficient Methods

I design optimization objectives and operator-theoretic techniques that trade off accuracy, compute, and reliability in a principled way. The goal is to deliver models that run fast under limited resources while retaining checkable guarantees about stability and performance, so they can be deployed confidently in time-sensitive settings.

5. Applications: Climate Science & Natural Hazards

I apply the above methods to geospatial and environmental data to support tasks such as flood and storm nowcasting, exposure mapping, and coastal risk assessment. The emphasis is on transparent, auditable models that policymakers and engineers can trust, turning advanced mathematics into practical tools for risk and resilience.

Let's Collaborate

I welcome collaborations on interpretable AI, as well as its applications in climate and natural hazards. 📧 csplevuanh@gmail.com · anhlv@ioit.ac.vn

Thanks for visiting. "In God we trust. All others must bring data."