Abstract

New advances in theories of artificial intelligence have returned to tackling the classical problem of reconciling symbolic reasoning and probabilistic statistical modelling of data. Methods in symbolic AI highlight representational power compositional structures and logical reasoning providing both interpretability and strong guarantees for reasoning tasks. Statistical models of data such as those used in deep learning models are very powerful at performing robust pattern recognition but do not provide interpretability and are generally not compositional. Recent advances in theories of AI seek to unite the two approaches by building neuro-symbolic systems that combine the generalisation abilities and efficient optimisation algorithms of statistical models of data with the formal structures provided by logic systems. Such challenges include causally-grounded generalisation compositional rule learning and verifiably robust and interpretable system design. Developments in these fields w