Reproduce “Task Programming: Learning Data Efficient Behavior Representations” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jennifer J. Sun, Ann Kennedy, Eric Zhan, David J. Anderson, Yisong Yue, Pietro Perona
The research from Caltech and Northwestern University introduces Trajectory Embedding for Behavior Analysis (TREBA), a framework that leverages expert-defined 'task programs' to guide self-supervised learning of trajectory representations. This method reduces the required data annotation for behavior classification by factors ranging from 2x to 20x across mouse and fly datasets, while maintaining or improving performance compared to fully supervised baselines.
Key insights
- Structured expert knowledge, when explicitly encoded via 'task programming' and integrated into self-supervised learning, significantly enhances the data efficiency for behavior classification models.
- Learning representations that directly encode expert-defined attributes (e.g., via attribute decoding or contrastive loss) is more effective than solely enforcing attribute consistency on generated trajectories.
- A small, diverse set of expert-engineered programs can effectively guide representation learning, achieving performance comparable to using a larger set of programs while reducing the variance of individual program impact.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.