Identification of individual animals from position, orientation, and movement sequences in the Morris water task with machine learning: implications for spatial learning.

The Morris water task (MWT) has been the gold standard assessment of place learning and memory for over four decades, yet the processes that contribute to successful navigation are not completely understood. It is generally posited that rats learn to navigate directly to a place; however, navigation is not perfectly direct but contains variation in swim speed and trajectory, resulting in efficient, yet suboptimal paths. Recently, our lab demonstrated that path topography is quantitatively similar across trials within individual rats for a given start location, and dissimilar between rats and different start locations within rat. This suggests rats may take unique paths to the platform from each release point. If the sequences of position and heading changes in the MWT are reinforced based on individual experience, it should be possible to objectively classify (identify) individual rats using machine learning. In the present study, a recurrent neural network was trained to classify 15 rats based on trial-level time-series data of path coordinates, velocity, and orientation. Testing on unseen coordinates or combined features from the same starting locations yielded classification accuracies that were 5.93- to 7.26-fold greater than chance and significantly greater than data from different start locations. Classification using scalar metrics with support vector machines was unable to match these outcomes. Collectively, these observations support the hypothesis that navigation in the MWT involves learning to perform behavioral sequences.