Feasibility and extent of strong integration of complex formal logics with deep learning

Determine whether a much stronger integration of complex formal logics and deep learning—incorporating best-of-both-worlds features such as explainability and provable correctness on the symbolic side and trainability from raw data and robustness on the neural side—is achievable, and ascertain the extent to which such integration can be realized.

Background

The paper surveys recent neuro-symbolic AI research and categorizes contemporary approaches, noting increased use of complex formal logics relative to earlier eras. Despite progress, many systems still operate primarily at the non-logical symbolic level, and commonly anticipated benefits of neuro-symbolic integration—such as interpretability and robust generalization—are not consistently realized.

In the forward-looking discussion, the authors emphasize the need to deepen logical aspects within neuro-symbolic AI and to build a systematic toolbox for leveraging complex logics alongside deep learning. They highlight that the core question of whether and to what extent a stronger integration is possible remains unresolved and fundamental to advancing the field.