Controllable AI: Control Theory meets Artificial Intelligence

Control theory is fundamental in the design and understanding of many natural and engineered systems, from cars and robots to power networks and bacterial metabolism. In this talk, we explore how the principles of control and dynamical systems ??"formalized with control theory??" can also play an important role in enhancing Artificial Intelligence (AI). We argue that AI systems are themselves dynamical in nature, and that meaningful dynamics emerge at multiple levels of granularity: from individual computational units that mix information, to full deep layered architectures, and ultimately to the real-world systems in which AI is deployed (such as embodied intelligence). We discuss two specific examples of different levels of granularity where ideas from control and dynamical systems lead to practical advances: first, how analyzing individual information-mixing modules enables improved architecture design grounded in dynamical principles; and second, how applying control-theoretic tools at the level of deep layered architectures allows us to steer and constrain model behavior with formal guarantees. Lastly, we give an overview of how examining AI across these different scales of granularity through a unified control lens reveals new opportunities for principled design, analysis, and control, ultimately moving us toward more efficient, predictable, and controllable AI systems. The aim of this talk is to illustrate the potential of viewing AI through the tools of control and dynamical systems, and to open a discussion about future research directions at the intersection of learning, control, and intelligence.

Speaker: Carmen Amo Alonso, Stanford University