Integration of Logical Constraints in Deep Learning

Track Editors

Alessandro Abate, University of Oxford, U.K.
Eleonora Giunchiglia, Imperial College London, U.K.
Bettina Könighofer, Graz University of Technology, Austria
Luca Pasa, University of Padova, Italy
Matteo Zavatteri, University of Padova, Italy

Overview

Over the last few years, the integration of logical constraints in Deep Learning models has gained significant attention from research communities for its potential to enhance the interpretability, robustness, safety, and generalization capabilities of these models. This integration opens the possibility of incorporating prior knowledge, handling incomplete data, and combining symbolic and subsymbolic reasoning. Moreover, the use of logical constraints improves generalization, formal verification, and ethical decision-making. The versatility of logical constraint integration spans diverse domains, presenting both research challenges and opportunities.

In recent times, there has been a growing trend in incorporating logical constraints into deep learning models, especially in safety-critical applications. Looking ahead, challenges in this field extend to the development of Machine Learning models that not only incorporate logical constraints but also provide robust assurances. This involves ensuring that AI systems adhere to specific (temporal) logical or ethical constraints, offering a level of guarantees in their behavior.

Thus, this special track seeks submissions on the integration of logical constraints into deep learning approaches. We are particularly interested in the following broad content areas.

  • Formal verification of neural networks is an active area of research that has been proposing methods, tools, specification languages (e.g., VNNLIB), and annual competitions (e.g., VNN-COMP) devoted to verify that a neural network satisfies a certain property typically given in (a fragment) of first order logic.
  • Synthesis aims at synthesizing neural networks that are compliant with some given constraint. Approaches to achieve this aim range from modifying the loss function in the training phase (i.e., soft constraint injection) to exploit counterexample guided inductive synthesis (CEGIS).
  • Monitoring: Logical constraints can be used to mitigate and/or neutralize constraint violations of machine learning systems when formal verification and synthesis are not possible. Shielding techniques intervene by changing the output of the network when a constraint is being violated. Runtime monitoring can be used to anticipate failures of AI systems without modifying them.
  • Explainability: Automated learning of formulae and logical constraints from past executions of the system provides natural explanations for neural network predictions and poses another avenue for future research. Formulae and constraints offer a high degree of explainability since they carry a precise syntax and semantics, and thus they can be "read" by humans more easily than other explainability methods.

This special track aims to explore and showcase recent advancements in the integration of logical constraints within deep learning models, spanning the spectrum of verification, synthesis, monitoring and explainability, by considering exact and approximate solutions, online and offline approaches. The focus will also extend to encompass innovative approaches that address the challenges associated with handling logical constraints in neural networks.

Call for Submissions

This special track seeks contributions that delve into various aspects of logic constraint integration in deep learning, including, but not limited to:

  • Learning with logical constraints
  • Enhancing neural network expressiveness for logical constraints
  • Formal verification (certification) of neural networks
  • Automated synthesis of certified neural networks, or of AI systems with neural nets
  • Decision making: Strategy/policy synthesis for AI systems with neural networks
  • Runtime monitoring of AI systems
  • Learning of (temporal) logic formulae for explainable and interpretable AI
  • Scalability challenges in neural networks with logical constraints
  • Real-world applications of neural networks with logical constraints
  • Enhancing model explainability via logical constraints
  • Design of neural networks under temporal logical requirements

Pertinent review papers of exceptional quality may also be considered.

Prospective authors should read the JAIR Policy on Special Tracks and the information on JAIR Submissions.

Key Dates

JAIR special tracks have a submission window. Papers can be submitted anytime during that window. They are reviewed as they arrive and accepted papers will go to production and will be published asynchronously as soon as they are ready, first as part of a usual JAIR pipeline and later on a JAIR webpage dedicated to the special track.

Target timeline:

  • Submission period: December 15, 2024 - May 31, 2025
  • First round of review and authors' notification: September 2025
  • Resubmissions: November 2025
  • Second round of review and authors' notification: January 2026
  • Final manuscripts: March 2026