Deep learning-based decoding of spatial information from limbic-cortical local field potentials reveal drifting spatial representations with increasing stability

Neural representations of space in the hippocampus and related brain areas change over timescales of days-weeks, even when there are no apparent behavioural changes. This `representational drift' occurs even after animals are fully familiar with a given context. Many qualities of this phenomenon are unknown, yet few tools exist to aid analysis. Here we present a novel deep-learning approach for robust quantification and analysis of ensemble level representational drift. Using this method, we analyse a longitudinal dataset of 0.5-475Hz broadband local field potential (LFP) data taken from Hippocampal, Prefrontal-Cortex and Parietal-Cortex of rats collected over multiple days, before and after a contextual rule change in a spatial navigation learning task. First, we observed clear spatial representations in all considered brain regions, despite the low frequency LFP data used. Second, we show statistically significant drift in these representations in all brain regions. Lastly, we show a statistically significant increase in the stability of representations for all considered brain regions as time and experience increases. Our general strategy for using deep neural networks to quantify drift in broadband LFP data opens up new possibilities for flexibly dissecting the features of drift in large-scale neural recordings, and how they relate to animal behaviour.