https://www.dagstuhl.de/22291
July 17 – 22 , 2022, Dagstuhl Seminar 22291
Machine Learning and Logical Reasoning: The New Frontier
Organizers
Sébastien Bardin (CEA LIST, FR)
Vijay Ganesh (University of Waterloo, CA)
Somesh Jha (University of Wisconsin-Madison, US)
Joao Marques-Silva (CNRS – Toulouse, FR)
For support, please contact
Jutka Gasiorowski for administrative matters
Michael Gerke for scientific matters
Documents
List of Participants
Shared Documents
Dagstuhl Seminar Wiki
Dagstuhl Seminar Schedule (Upload here)
(Use personal credentials as created in DOOR to log in)
Motivation
Machine learning (ML) and logical reasoning have been the two key pillars of AI since its inception, and yet, there has been little interaction between these two sub-fields over the years. At the same time, each of them has been very influential in their own way. ML has revolutionized many sub-fields of AI including image recognition, language translation, and game playing, to name just a few. Independently, the field of logical reasoning (e.g., SAT/SMT/CP/first-order solvers and knowledge representation) has been equally impactful in many contexts in software engineering, verification, security, AI, and mathematics. Despite this progress, there are new problems, as well as opportunities, on the horizon that seem solvable only via a combination of ML and logic.
One such problem that requires one to consider combinations of logic and ML is the question of reliability, robustness, and security of ML models. For example, in recent years, many adversarial attacks against ML models have been developed, demonstrating their extraordinary brittleness. How can we leverage logic-based methods to analyze such ML systems with the aim of ensuring their reliability and security? What kind of logical language do we use to specify properties of ML models? How can we ensure that ML models are explainable and interpretable?
In the reverse direction, ML methods have already been successfully applied to making solvers more efficient. In particular, solvers can be modeled as complex combinations of proof systems and ML optimization methods, wherein ML-based heuristics are used to optimally select and sequence proof rules. How can we further deepen this connection between solvers and ML? Can we develop tools that automatically construct proofs for higher math?
To answer these and related questions we are organizing this Dagstuhl Seminar, with the aim of bringing together the many world-leading scientists who are conducting pioneering research at the intersection of logical reasoning and ML, enabling development of novel solutions to problems deemed impossible otherwise.
Motivation text license Creative Commons BY 4.0
Sébastien Bardin, Vijay Ganesh, Somesh Jha, and Joao Marques-Silva
Dagstuhl Seminar Series
- 19062: "Bringing CP, SAT and SMT together: Next Challenges in Constraint Solving" (2019)
- 14442: "Symbolic Execution and Constraint Solving" (2014)
Classification
- Logic In Computer Science
- Machine Learning
- Software Engineering
Keywords
- SAT/SMT/CP solvers and theorem provers
- Testing
- Analysis
- Verification
- And Security of Machine Learning
- Neuro-symbolic AI