Hosted by the Alan Turing Institute's Neuro Symbolic AI Interest Group
Abstract: Many perceive neural and symbolic computation to be fundamentally incompatible due to the belief that symbolic computation ‘is not differentiable’. I beg to differ: Many logics are indeed differentiable, and we can seamlessly integrate them with neural networks to create end-to-end differentiable systems. Practical implementations, however, require understanding how these integrations work in real-world contexts. We discuss several challenges: Unexpected reasoning semantics, difficulties in scaling up to large-scale reasoning, and limited support for current generative AI models. For each challenge, we explore potential solutions and opportunities for improved understanding.
Join online: https://ed-ac-uk.zoom.us/j/83495654059