To reduce annotation effort, we present TREBA: a method to learn annotation-sample efficient trajectory embedding for behavior analysis, based on multi-task self-supervised learning. This issue arises prominently in automated behavior analysis, in which agent movements or actions of interest are detected from video tracking data. ![]() Specialized domain knowledge is often necessary to accurately annotate training sets for in-depth analysis, but can be burdensome and time-consuming to acquire from domain experts. Our results thus suggest that task programming and self-supervision can be an effective way to reduce annotation effort for domain experts. Using embeddings from TREBA, we reduce annotation burden by up to a factor of 10 without compromising accuracy compared to state-of-the-art features. We present experimental results in three datasets across two domains: mice and fruit flies. We evaluate this trade-off using data from behavioral neuroscience, in which specialized domain knowledge is used to identify behaviors. Total domain expert effort can be reduced by exchanging data annotation time for the construction of a small number of programmed tasks. The tasks in our method can be efficiently engineered by domain experts through a process we call "task programming", which uses programs to explicitly encode structured knowledge from domain experts. Looking out into the future, I will conclude by highlight selected open questions and challenges that have high potential for impact in video games and raise key questions for current state-of-the-art AI approaches. This competition is built around a complex task, large-scale demonstration data, and an evaluation setup that requires and rewards sample efficient learning and effective generalization. Sample efficient learning is a key challenge, with current algorithms often requiring millions of samples to learn to perform individual narrow tasks, limiting the scope and applicability of these approaches. This ambitious competition is designed to drive advances in sample efficient reinforcement learning with human priors. I will highlight our most recent collaboration, led by a team of PhD students at Carnegie Mellon University: the MineRL competition. Here, I will illustrate the capabilities of the platform with recent examples that I find particularly exciting. To unlock the potential of Minecraft for AI experimentation, my team has developed Project Malmo - an open source experimentation platform built on top of Minecraft to enable a wide range of research. This open-ended nature both make the game appealing to its human fan-base, and uniquely challenging to AI algorithms. Minecraft is an open-world game, where players explore, create, and continuously find new ways to play and engage with each other. This talk focuses on opportunities in the setting of the game Minecraft, one of the most popular video games of all time. If and when these are successfully tackled, new algorithms and insights have the potential to enable entirely new game experiences. For driving research, games provide rich data that can be used to tackle hard problems, from complex decision making to collaboration. Modern video games provide exciting challenges and opportunities for pushing the state of the art in machine learning and other research areas, and, in turn, stand to first benefit from research advances.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |