Deep learning-based location decoding reveals that across-day representational drift is better predicted by rewarded experience than time

Neural representations of locational and spatial relations in the hippocampus and related brain areas change over timescales of days-weeks, even in familiar contexts and when behavior appears stable. It remains unclear how this representational drift is driven by combinations of the passage of time, general experience or specific features of experience. We present a novel deep-learning approach for measuring network-level representational drift, quantifying drift as the rate of change in decoder error of deep neural networks as a function of train-test lag. Using this method, we analyse a longitudinal dataset of 0.5-475 Hz broadband local field potential (LFP) data recorded from dorsal hippocampal CA1, medial prefrontal cortex and parietal cortex of six rats over [~]30 days, during learning of a spatial navigation task in an initially unfamiliar environment. All three brain regions contained clear spatial representations which evolve and drift over training sessions. We find that the rate of drift slows for later training sessions. Finally, we find that drift is statistically better explained by task-relevant experiences within the maze, rather than the passage of time or number of sessions the animal spent on the maze. While previous research has focused on drift as a measure of the changes in spiking activities of units, here we examine drift as a measure of change in oscillatory activity of local field potentials; our approach of using of deep neural networks to quantify drift in broadband neural time series unlocks new possibilities for defining the drivers and functional consequences of representational drift.