A predictive role of the old cortex in general intelligence

Cross-disciplinary research has revealed many parallels between neocortical sensory information processing and that found in artificial neural networks. Here, we argue that neural and computational insights into the hippocampus, a structurally unique and phylogenetically old cortical region, have yet to be fully leveraged by such models. The hippocampus contains a well-described neural circuitry and is situated at the nexus of multiple brain networks, allowing it to maintain latent sensory, affective, and higher order transmodal information. Recent AI developments based on self-supervised architectures that engage prediction in latent spaces are data-efficient and obey real world constraints, making them promising not only as AI models but also as neuroscientific models. Within that framework, the hippocampus can be seen as computing and maintaining predictions over multiple time spans, thereby increasing the number of comparisons between predicted and real latent sensory data. This may help explain why animals with a hippocampus are able to learn so efficiently compared to popular, data-hungry AI architectures. Our perspective offers new avenues towards understanding neural intelligence and the design of data-efficient, naturalistic AI.