Skip to contentSkip to filters

We use cookies for analytics.

MABe22: A Multi-Species Multi-Task Benchmark for Learned Representations of Behavior

Abstract

Jennifer J. Sun and Markus Marks present MABe22, a benchmark for testing what learned representations preserve about interacting animals. It combines mouse video and pose tracking, footage of beetle–ant interactions and fly trajectories. Evaluation tasks test whether representations retain information about behavior and experimental conditions. The authors compare self-supervised video and trajectory methods and find that approaches developed for human action data transfer imperfectly across species and settings.

Topics