Certified Learning
Ensuring hard constraints in machine learning models.
Learning with Projection Layers
A first line of my work develops neural controllers with embedded optimization layers that enforce safety constraints during learning and deployment. In Opt-ODENet, we integrate differentiable quadratic programming layers with Neural ODE controllers, so that control barrier function constraints can be imposed as hard requirements while the controller is trained end-to-end. Stability is promoted through control Lyapunov function conditions in the learning objective, while safety is enforced by the projection layer in real time. This provides a way to combine the expressive power of neural controllers with certificate-based safety mechanisms.
Learning with Certified Parameterizations
A second line of my work focuses on building structural guarantees directly into the learning architecture. In learning Koopman representations with controllability guarantees, the goal is not only to fit a predictive latent model, but also to ensure that the learned representation preserves a key control-theoretic property: controllability. By parameterizing the Koopman surrogate in controllability-preserving canonical forms, the learned model becomes compatible with downstream control design methods such as MPC. This approach shifts certification from post-hoc verification to by-construction learning.
Selected Publications
Learning with Neural Certificates and Verification
A third line of my work studies how neural networks can be trained together with formal certificates of stability and optimality. In learning neural controllers with input-output dissipativity, we use neural storage functions and supply rate functions to certify closed-loop behavior. Since dissipativity generalizes Lyapunov stability and connects naturally to optimal control, this framework allows neural controllers to inherit both stability guarantees and performance-oriented optimality properties. This direction aims to make neural control not only expressive, but also verifiable and certifiable.