We are facing a quantum leap in scalable approaches to Machine Intelligence (MI), but there seems to be a gap between (a) the impressive ability of MI to mimic predictive interaction and contextual adaptiveness, and (b) its substantial ignorance about the reasons why that ability concerns anything humans are interested in. The gap can be described as a lack of grounding in the compositional features of AI, be them symbolic or sub-symbolic, uni- or multi-modal, informed by background knowledge and representational learning, or not.
Professor Gangemi will present some challenges and partial solutions to grounded compositionality in AI, and will discuss a few positions on the matter of AI evaluation, based on mixed technical and foundational issues.