Neurosymbolic Control for Verifiable Design of AI-Enabled Systems

Motivation

The availability of data and the advances in machine/robot learning and artificial intelligence (AI) have resulted in great progress in the design of highly performant systems, such as self-driving cars, mobile robots, and transportation systems. Indeed, much progress was made in the design of AI components, e.g., for perception tasks, trajectory prediction, decision-making, and low-level nonlinear control. However, the integration of these AI components into safety-critical control systems has mostly been ad-hoc and without correctness guarantees which has resulted in unsafe system behavior, e.g., due lacking robustness of object detectors or jailbroken LLM-controlled robots. 

In contrast, the focus in more traditional domains such as systems & control theory, formal methods, and optimization has been on system design that enjoys rigorous and verifiable correctness guarantees. To name a few examples,  algorithms that provide strong safety and performance guarantees include Hamilton-Jacobi reachability analysis, symbolic control synthesis, game theory, and optimization-based control. Obtaining such strong control guarantees typically comes at the cost of  increased computational complexity, making these algorithms computationally intractable for higher-dimensional systems (such as AI-enabled or multi-agent systems) and more complex system requirements (such as temporal logic constraints). 

On the one hand, we thus have efficient and highly performant but hard to verify AI-enabled algorithms, while, on the other hand, we have verifiably correct but often computationally intractable algorithms. To bridge this gap, in this workshop, we investigate and discuss novel concepts, theories, and algorithms for neurosymbolic systems that combine highly performing AI models (such as neural network representations) with verifiable rule-based processing techniques (also referred to as symbolic methods) to improve safety, explainability, and performance of AI systems. Roughly, there are three levels of abstraction for neurosymbolic approaches: (1) a neural implementation of a logic, (2) a logical characterization of a neural system, and (3) a hybrid system combining a neural model and a logical characterization. In neurosymbolic systems we hence study neural network models and logical reasoning techniques as integrated models of computation. While the advancements in neurosymbolic systems present exciting opportunities towards building efficient and performant methods for AI systems, it also raises new questions and introduces novel challenges. Our workshop on "Neurosymbolic Control for Verifiable Design of AI-Enabled Systems" aims to unravel these challenges and open problems for a general audience and discuss what new principles and techniques we need. For this purpose, we have invited nine renown expert speakers (find the speaker info below).

Speakers and Panelists

Tentative Schedule

Organizers