In this study, researchers used large-scale, longitudinal two-photon calcium imaging to observe the formation of cognitive maps in the hippocampus of mice. The mice learned to navigate through virtual reality tracks and collect rewards. The results showed that both the behavior of the mice and the neural activity in the hippocampus progressed through various stages, leading to improved task understanding and efficiency. The neural activity also displayed a state machine-like structure, which was captured by a Hidden Markov Model and a recurrent neural network trained using Hebbian learning. Interestingly, sequence models like LSTMs and Transformers did not naturally produce similar representations. The findings provide insights into the mathematical form of cognitive maps and the learning processes involved, contributing to our understanding of biological intelligence and potential applications in artificial intelligence.
https://www.biorxiv.org/content/10.1101/2023.08.03.551900v2