Neuro-Symbolic Learning of Answer Set Programs from Raw Data

• DOI: 10.24963/ijcai.2023/399
• Publication Date: 2023-08-11T04:31:30Z
• Authors: Daniel Cunnington; Mark Law; Jorge Lobo; Alessandra Russo
• Keywords: Interpretability; Inductive bias

One of the ultimate goals Artificial Intelligence is to assist humans in complex decision making. A promising direction for achieving this goal Neuro-Symbolic AI, which aims combine interpretability symbolic techniques with ability deep learning learn from raw data. However, most current approaches require manually engineered knowledge, and where end-to-end training considered, such are either restricted definite programs, or binary neural networks. In paper, we introduce Inductive Learner (NSIL), an approach that trains a general network extract latent concepts data, whilst knowledge maps target labels. The novelty our method biasing based on in-training performance both components. We evaluate NSIL three problem domains different complexity, including NP-complete problem. Our results demonstrate learns expressive solves computationally problems, achieves state-of-the-art terms accuracy data efficiency. Code technical appendix: https://github.com/DanCunnington/NSIL