Reliable and Safe Machine Learning via Built-in Hard Constraints

Abstract

ML models are increasingly being deployed in domains where respecting safety, physical, or structural constraints is critical. Yet most modern learning systems encode such requirements through data or soft penalties, providing little assurance when deployed beyond the training distribution. In this talk, I will discuss recent efforts by my group toward building hard constraints directly into learning-based decision-making systems, while retaining the flexibility and scalability that make modern machine learning effective. I will begin with hard-constrained neural networks (HardNets), which add a differentiable constraint-enforcement layer into existing architectures, enabling end-to-end learning under various constraints while preserving expressivity and guaranteeing feasibility. I will then discuss how HardNets can be used to address important problems in dynamical systems and control, including learning chaotic dynamics with boundedness guarantees via energy-constrained operators, as well as safe yet scalable learning-based control and reinforcement learning. Finally, I will turn to modern generative models, namely diffusion and flow-matching models, where treating sampling as an optimal control problem allows us to impose constraints on the final generated sample without unnecessarily restricting the intermediate sampling dynamics, with applications spanning robotic manipulation and planning, PDE control, and language-guided image editing. Together, these advances pave the way for reliable decision-making systems with intrinsic guarantees.

Related papers

• HardNet: Hard-Constrained Neural Networks with Universal Approximation Guarantees
• HardNet++: Nonlinear Constraint Enforcement in Neural Networks
• ECO: Energy-Constrained Operator Learning for Chaotic Dynamics with Boundedness Guarantees
• PS2-RL: Provably Safe, Yet Scalable Reinforcement Learning
• HardFlow: Hard-Constrained Sampling for Flow-Matching Models via Trajectory Optimization
• DiRecT: Safe Diffusion-Based Planning via Receding-Horizon Denoising

Bio

Navid Azizan is the Alfred H. (1929) and Jean M. Hayes Associate Professor at MIT, where he is a Principal Investigator in the Laboratory for Information & Decision Systems (LIDS) and holds dual appointments in the Department of Mechanical Engineering (in Control, Instrumentation & Robotics) and the Schwarzman College of Computing’s Institute for Data, Systems & Society (IDSS). His research interests broadly lie in machine learning, systems and control, and mathematical optimization. His research lab focuses on various aspects of reliable AI systems, with applications to high-stakes and safety-critical settings. He obtained his PhD in Computing and Mathematical Sciences (CMS) from the California Institute of Technology (Caltech) in 2020, his MSc in electrical engineering from the University of Southern California in 2015, and his BSc in electrical engineering and physics from Sharif University of Technology in 2013. Prior to joining MIT, he completed a postdoc at Stanford University in 2021. Additionally, he was a research scientist intern at Google DeepMind in 2019. He is a recipient of several awards, including the National Science Foundation CAREER Award, research awards from Amazon, Google, and MathWorks, among others, and the inaugural Information Theory and Applications (ITA) “Sun” (Gold) Graduation Award. His work has been recognized with best paper awards at several venues, including Learning for Dynamics and Control (L4DC), the INFORMS Junior Faculty Paper Competition, and ACM Greenmetrics. He was named in the list of Outstanding Academic Leaders in Data by the CDO Magazine for two consecutive years in 2024 and 2023. His teaching and mentorship have been recognized with the Joseph A. Martore (1975) Excellence in Teaching Award, the Frank E. Perkins Award for Excellence in Graduate Advising (MIT Institute Award), and the UROP Outstanding Mentor Award.