Neurosymbolic AI Combines Answer Set Programming with Energy Models

A new neurosymbolic AI methodology has been developed that seamlessly integrates Answer Set Programming (ASP) with energy-based models. This innovative approach aims to combine the strengths of symbolic reasoning with continuous learning, offering a powerful platform for complex AI tasks. Key contributions include enabling joint optimization within continuous latent spaces, where explicit ASP-based declarative semantics fully incorporate background knowledge, constraints, and non-monotonic inference. This allows the system to leverage robust logical reasoning alongside probabilistic learning. The methodology also advances existing work by providing a generalized model and a practical platform for ASP-centric, end-to-end training, particularly for applications in dynamic domains that involve perception and interaction. Demonstrations with MNIST, visual question-answering (Clevr), and multi-object tracking (MOT) benchmarks illustrate its basic use and application.