Dagstuhl-Seminar 19361
Logic and Learning
( 01. Sep – 06. Sep, 2019 )
(zum Vergrößern in der Bildmitte klicken)
Permalink
Bitte benutzen Sie folgende Kurz-Url zum Verlinken dieser Seite:
https://www.dagstuhl.de/19361
Organisatoren
- Michael Benedikt (University of Oxford, GB)
- Kristian Kersting (TU Darmstadt, DE)
- Phokion G. Kolaitis (University of California - Santa Cruz & IBM Almaden Research Center - San Jose, US)
- Daniel Neider (MPI-SWS - Kaiserslautern, DE)
Kontakt
- Andreas Dolzmann (für wissenschaftliche Fragen)
- Jutka Gasiorowski (für administrative Fragen)
Gemeinsame Dokumente
- Dagstuhl Materials Page (Use personal credentials as created in DOOR to log in)
Impacts
- Conjunctive Queries : Unique Characterizations and Exact Learnability : article - Cate, Balder D. ten; Dalmau, Victor - Cornell University : arXiv.org, 2020. - 29 pp..
- Graph Neural Networks Meet Neural-Symbolic Computing: A Survey and Perspective : article in IJCAI-20 - Lamb, Luis C.; D'Avila Garcez, Artur S.; Gori, Marco; Prates, Marcelo O. R.; Avelar, Pedro H. C.; Vardi, Moshe Y. - IJCAI, 2020. - 8 pp..
- Learning Properties in LTL ∩ ACTL from Positive Examples Only : article - Wien : Universität, 2020. - 9 pp. - Ehlers, Rüdiger; Gavran, Ivan; Neider, Daniel - Wien : Universität, 2020. - 9 pp..
- The Homomorphism Lattice, Unique Characterizations, and Concept Learning : article - Cate, Balder D. ten - Aachen : CEUR, 2020. - 9 pp..
Programm
Logic and learning are central to Computer Science, and in particular to AI-related research. Already Alan Turing envisioned in his 1950 "Computing Machinery and Intelligence" paper a combination of statistical (ab initio) machine learning and an "unemotional" symbolic language such as logic. Currently, however, research in logic and research in learning interact far too little with each other; in fact, they are often perceived as being completely distinct or even opposing approaches. While there has been interest in using machine learning methods within many application areas of logic, the investigation of these interactions has usually been carried out within the confines of a single problem area. We believe that an interaction involving a broader perspective is needed. It would be fruitful to look for common techniques in applying learning to logic-related tasks, which requires looking across a wide spectrum of applications. It is also important to consider the ways that logic and learning, deduction and induction, can work together.
The main aim of this Dagstuhl Seminar is to bring together researchers from various communities related to logic and learning, and to create bridges between the two fields via the exchange of ideas ranging from the injection of declarative methods in machine learning to uses and applications of learning in logical contexts. This will include creating an understanding of the work in different applications, as well as an increased understanding of the formal connections between these applications and the development of a more unified view of the current attempts to synthesize deductive and inductive approaches. The seminar will explore the following three distinct strands of interaction between logic and learning.
- Machine Learning for Logic, including the learning of logical artifacts, such as formulas, logic programs, database queries and integrity constraints, as well as the application of learning to tune deductive systems.
- Logic for Machine Learning, including the role of logics in delineating the boundary between tractable and intractable learning problems, the construction of formalisms that allow learning systems to take advantage of specified logical rules, and the use of logic as a declarative framework for expressing machine-learning constructs.
- Logic vs. Machine Learning, including the study of problems that can be solved using either logic-based techniques or via machine learning, the exploration of the trade-offs between adopting logic-based methods vs. adopting learning-based methods in cases where both methods apply, and the development of benchmarks for comparing these methods.
Creative Commons BY 3.0 DE
Michael Benedikt, Kristian Kersting, Phokion G. Kolaitis, and Daniel Neider
Motivation
Logic and learning are central to Computer Science, and in particular to AI research and allied areas. Alan Turing envisioned, in his paper "Computing Machinery and Intelligence" [1], a combination of statistical (ab initio) machine learning and an "unemotional" symbolic language such as logic. However, currently, the interaction between research in logic and research in learning is far too limited; in fact, they are often perceived as being completely distinct or even opposing approaches.
While there has been interest in using machine learning methods within many application areas of logic, the investigation of these interactions has usually been carried out within the confines of a single problem area. We believe that an interaction involving a broader perspective is needed. It would be fruitful to look for common techniques in applying learning to logic-related tasks, which requires looking across a wide spectrum of applications. It is also important to consider the ways that logic and learning, deduction and induction, can work together.
Design of the Seminar
The main aim of this Dagstuhl Seminar was to address the above problems by bring researchers from the logic and learning communities together and to create bridges between the two fields via the exchange of ideas ranging between the (seemingly) polar possibilities of the injection of declarative methods in machine learning and the use and applications of learning technologies in logical contexts. This included creating an understanding of the work in different applications, an increased understanding of the formal connections between these applications, and the development of a more unified view of the current attempts to organically reconcile deductive and inductive approaches. In order to structure these explorations, the focal points of the seminar were the following three distinct strands of interaction between logic and learning:
- Machine Learning for Logic, including the learning of logical artifacts, such as formulas, logic programs, database queries and integrity constraints, as well as the application of learning to tune deductive systems.
- Logic for Machine Learning, including the role of logics in delineating the boundary between tractable and intractable learning problems, the construction of formalisms that allow learning systems to take advantage of specified logical rules, and the use of logic as a declarative framework for expressing machine learning constructs.
- Logic vs. Machine Learning, including the study of problems that can be solved using either logic-based techniques or via machine learning, an exploration of the trade-offs between these techniques, and the development of benchmarks for comparing these methods.
Summary of seminar activities
The seminar was attended by 41 researchers across various communities including logic, databases, Inductive Logic Programming (ILP), formal verification, machine learning, deep learning, and theorem proving. The membership consisted of senior and junior researchers, including graduate students, post-doctoral researchers, and industry experts. The seminar was conducted through talks and breakout sessions, with breaks for discussion between the attendees. There were three long talks, 21 short talks, and three breakout sessions on the discussion of open problems in logic and learning.
The talks consisted of: (i) presentation of recent advances in research questions and methodologies relating to the motivations discussed above; (ii) surveys of the state of research on various problems requiring the combination of deductive and inductive reasoning as well as methodologies developed to address fundamental hurdles in this space; (iii) new perspectives on the organic combination of logical formulations and methods with machine learning in specific application domains; (iv) theoretical formulations and results on problems in learning logical representations; (v) demonstrations of state-of-the art tools combining logic and learning for applications such as theorem proving or entity resolution; (vi) presentation of research on challenge problems for the field of AI and intelligent reasoning.
The breakout sessions were conducted in three continuing parts, each spanning one session. The first part involved all the participants in a discussion of the current (small and large) open problems in AI, challenge problems for the field of intelligent systems, and research questions about defining specific goals representing a successful combination of inductive and deductive reasoning. This involved a deliberation of what problems were relevant, which problems could be potentially related to or dependent upon each other, and various suggestions to formalise commonly desired research goals. This session resulted in the choice of three broad areas for further specific discussion: (i) Explainable AI (ii) Injecting symbolic knowledge or constraints into neural networks, and (iii) Learning of logical formulae (first-order logic) from satisfaction on structures in a differentiable manner. The second part consisted of parallel thematic sessions on these three areas. Each thematic session was conducted in the form of a round-table discussion and was led by one or two participants who championed the theme. The third session brought all the participants together again to conclude with a summary of the ideas exchanged during the parallel sessions.
Conclusion
We consider the seminar a success. There is a growing need to enable the disparate communities of logic and learning to interact with each other, and we noted from the seminar that researchers from each community appreciated the perspective offered by the other, often identified techniques used by the other community that could be imported into their own, and, interestingly, were in agreement about the relevant and important problems of the day. The format of the seminar including ample time for discussions and breakout sessions received positive feedback from the participants.
References
- A. M. Turing, “Computing machinery and intelligence”, Mind, vol. LIX, pp. 433–460, October 1950
Michael Benedikt, Kristian Kersting, Phokion G. Kolaitis, Adithya Murali, and Daniel Neider
- Isolde Adler (University of Leeds, GB) [dblp]
- Molham Aref (relational AI - Berkeley, US) [dblp]
- Vaishak Belle (University of Edinburgh, GB) [dblp]
- Michael Benedikt (University of Oxford, GB) [dblp]
- Ismail Ilkan Ceylan (University of Oxford, GB) [dblp]
- Victor Dalmau (UPF - Barcelona, ES) [dblp]
- Luc De Raedt (KU Leuven, BE) [dblp]
- Dana Fisman (Ben Gurion University - Beer Sheva, IL) [dblp]
- James Freitag (University of Illinois - Chicago, US) [dblp]
- Ivan Gavran (MPI-SWS - Kaiserslautern, DE) [dblp]
- Martin Grohe (RWTH Aachen, DE) [dblp]
- Barbara Hammer (Universität Bielefeld, DE) [dblp]
- Daniel Huang (University of California - Berkeley, US) [dblp]
- Nils Jansen (Radboud University Nijmegen, NL) [dblp]
- Brendan Juba (Washington University - St. Louis, US) [dblp]
- Kristian Kersting (TU Darmstadt, DE) [dblp]
- Sandra Kiefer (RWTH Aachen, DE) [dblp]
- Angelika Kimmig (Cardiff University, GB) [dblp]
- Phokion G. Kolaitis (University of California - Santa Cruz & IBM Almaden Research Center - San Jose, US) [dblp]
- Egor Kostylev (University of Oxford, GB) [dblp]
- Paul Krogmeier (University of Illinois - Urbana Champaign, US) [dblp]
- Luis C. Lamb (Federal University of Rio Grande do Sul, BR) [dblp]
- Carsten Lutz (Universität Bremen, DE) [dblp]
- Mateusz Malinowski (Google DeepMind - London, GB) [dblp]
- Henryk Michalewski (University of Warsaw, PL) [dblp]
- Adithya Murali (University of Illinois - Urbana-Champaign, US) [dblp]
- Sriraam Natarajan (University of Texas - Dallas, US) [dblp]
- Daniel Neider (MPI-SWS - Kaiserslautern, DE) [dblp]
- Dan Olteanu (University of Oxford, GB) [dblp]
- Ana Ozaki (Free University of Bozen-Bolzano, IT) [dblp]
- Madhusudan Parthasarathy (University of Illinois - Urbana-Champaign, US) [dblp]
- Lucian Popa (IBM Almaden Center - San Jose, US) [dblp]
- Martin Ritzert (RWTH Aachen, DE) [dblp]
- Xujie Si (University of Pennsylvania - Philadelphia, US) [dblp]
- Dan Suciu (University of Washington - Seattle, US) [dblp]
- Christian Szegedy (Google Inc. - Mountain View, US) [dblp]
- Balder Ten Cate (Google Inc. - Mountain View, US) [dblp]
- Josef Urban (Czech Technical University - Prague, CZ) [dblp]
- Steffen van Bergerem (RWTH Aachen, DE) [dblp]
- Guy Van den Broeck (UCLA, US) [dblp]
- Zsolt Zombori (Alfréd Rényi Institute of Mathematics - Budapest, HU) [dblp]
Klassifikation
- artificial intelligence / robotics
- data bases / information retrieval
- verification / logic
Schlagworte
- machine learning
- logic
- databases
- verification
- computational complexity
