Interpretable and Comparable Measurement for Computational Ethology

Abstract

Computational ethology, the field of automated behavioral analysis, has introduced a number of new techniques in recent years, including supervised, unsupervised, and statistical approaches for measuring behavior. These approaches have made the measurement of behavior more precise and efficient, improving significantly on human annotation of behavior. Despite this improvement, however, researchers in the field often interpret and report results primarily with reference to traditional, linguistic categories of behavior, which reduces measurement precision. I argue that using interpretable measurement techniques can help researchers to overcome this imprecision. Further, I present Basin, a system for performing interpretable behavioral measurement. Pioneering work has recently developed deep learning systems which can accurately estimate the pose of animals in videos. Basin builds on this work by creating a system in which researchers can construct behavioral measurements which are functions of an animal's pose. Specifically, Basin provides a graphical interface in which users can compose mathematical operations to construct behavioral measurements, without writing any code. Additionally, Basin provides a variety of features to visualize and refine behavioral measurements. I demonstrate that Basin can be used to replicate results from past studies of behavior and further show that Basin can be used to analyze the manifold structure of ant walking behavior.