JAABA: An interactive machine-learning tool for automatic annotation of animal behavior.

JAABA: An interactive machine-learning tool for automatic annotation of animal behavior. Kristin Branson¹, Mayank Kabra¹, Alice A. Robie¹, Marta Rivera-Alba^1,2, Steven Branson^1,3. 1) HHMI Janelia Farm Research Campus, Ashburn, VA; 2) Instituto Gulbenkian de Cięncia, Oeiras, Portugal; 3) Dept. of Computer Science and Engineering, UC San Diego, La Jolla, CA.

We present the Janelia Automatic Animal Behavior Annotator (JAABA), a new machine learning-based system to enable researchers to automatically compute interpretable, quantitative statistics describing video of behaving animals. Through our system, users encode their intuition about the structure of behavior by labeling the behavior of the animal, e.g. walking, grooming, or following, in a small set of video frames. JAABA uses machine learning techniques to convert these manual labels into behavior detectors that can then be used to automatically classify the behaviors of animals in large data sets with high throughput. JAABA combines an intuitive graphical user interface, a fast and powerful machine learning algorithm, and visualizations of the classifier into an interactive, usable system for creating automatic behavior detectors. We demonstrate that our system can be used by scientists without expertise in computer science to independently train accurate behavior detectors. Our system is general purpose, and can be used to easily create a wide variety of accurate individual and social behavior detectors for both adult and larvae Drosophila. We also show that it can be used to create behavior classifiers robust enough to successfully be applied to a large, phenotypically diverse data set consisting of thousands of transgenic lines of Drosophila melanogaster. Statistics of the automatic behavior classifications such as the fraction of time spent performing a given behavior are powerful descriptions, and we show that these statistics can be used to understand the subtle behavioral differences between highly similar populations of wildtype flies, between flies starved for differing amounts of times, and between flies of different ages. Our system is complementary to video-based tracking methods, and we envision that it will facilitate extraction of detailed, scientifically meaningful measurements of the behavioral effects in large experiments.