Machine Learning for the Quantification of Natural Mouse Behavior

Abstract

The study of cognitive disorders such as autism spectrum disorder is limited by our ability to measure and describe complex patterns of behavior. However, recent advances in computational ability and machine learning have enabled significant progress in computational behavior analysis in neuroscience.

One such method was developed by Berman et al. for the fruit fly Drosophila. By capturing raw video of a freely moving organism, performing some simple image processing and alignment, followed by principal component analysis to reduce dimensionality, a continuous wavelet transform to capture time-dependence, and t-distributed stochastic neighbor embedding for non-linear visualization, Berman et al. were able to produce a "behavior map" in an unsupervised manner where distinct regions correspond to visually distinguishable behavioral motifs.

We expand on the work of Berman et al. and Manley by addressing the specific engineering and scientific challenges of applying this methodology to study mice, a common model organism in neuroscience because of its more complex brain structure and behavioral repertoire. We also experiment with more advanced machine learning techniques in an attempt to better represent behavioral patterns for clustering. Finally, we demonstrate our methodology as a proof of principle of automated behavior quantification in mice by applying it to a cerebellar perturbation experiment.