Predictive Representations in the Hippocampal Formation

Recent breakthroughs in computational neuroscience and machine learning demonstrate
the brain can be understood as a predictive machine, refining its understanding of the world through future projections. Although influential, these models lack biological
mechanisms and empirical validation. Now experimental and theoretical advances are
revealing potential biological bases, providing methods for validation - bringing a
mechanistic neural-level understanding of how knowledge is extracted from
experience to guide behaviour within reach. This understanding could transform
fundamental neuroscience and catalyse neural-level understanding of cognitive symptoms of disorders like Schizophrenia, PTSD and Alzheimer's disease

A core insight is that neural networks can learn powerful latent variable representations by predicting their next inputs. These encapsulate how animals and artificial agents move between states in their environment or task - transition-structures - enabling prediction, planning and generalisation in new scenarios.

Grid cells, as the central layer of a generative network, learn the universal transition-structure of planar Euclidean environments - their activity resembling the eigenvectors of the transition matrix between states. Place cells represent environment-specific states - mapping grid-location to environmental features and capturing associative or predictive information from experience. Remarkably, analogous representations are observed in non-spatial tasks involving transitions within conceptual or relational spaces.

We will investigate how these structural representations are learned, with a particular focus on when different policies are necessary for adaptive behaviour. Our aim is to formulate and test a unified predictive model of neural circuits, neural firing-patterns, and neuromodulatory effects supporting behaviour in spatial tasks, focusing on two experimental streams.