Uploaded September 2025 | Updated September 2026, 1 week ago
Most AI systems trained on carefully curated datasets struggle when faced with the messy reality of naturalistic video. PooDLe [arxiv.org/abs/2408.11208], a newly published work, addresses this problem.
“We were interested in building embodied learning algorithms that could learn directly from video streams." Says Mengye Ren, author of the paper and assistant professor of computer science and data science at NYU.
The algorithm's name comes from a combination of a pooled loss function and a dense loss function used in PooDLe's architecture.
Visit nyu.edu/news for updates on this story and more.
Most AI systems trained on carefully curated datasets struggle when faced with the messy reality of naturalistic video. PooDLe [arxiv.org/abs/2408.11208], a newly published work, addresses this problem.
“We were interested in building embodied learning algorithms that could learn directly from video streams." Says Mengye Ren, author of the paper and assistant professor of computer science and data science at NYU.
The algorithm's name comes from a combination of a pooled loss function and a dense loss function used in PooDLe's architecture.
Visit nyu.edu/news for updates on this story and more.
