Improving the utility of synthetic vision data through search
Adriana Romero-Soriano - Meta FAIR
Feb. 27, 2026, 2:30 p.m. - Feb. 27, 2026, 3:30 p.m.
ENGMD 280
Hosted by: Oana Balmau
Over the last decade, the de facto standard for training high performing representation learning models has heavily relied on large scale static datasets crawled from the Internet. However, recent advances in visual content creation are challenging this status quo by pushing researchers to leverage high performing vision generative models as sources of training data. In this talk, I will address two research questions: (1) Are state-of-the-art image generative models optimized to provide high utility synthetic data?; and (2) How can we search the generative models’ manifold to find high utility synthetic data?
Adriana Romero-Soriano is a Research Scientist at Meta FAIR (Fundamental AI Research), an Adjunct Professor at McGill University's School of Computer Science, a core academic member of Mila (Quebec Artificial Intelligence Institute), and the recipient of a prestigious Canada CIFAR AI Chair. Adriana’s research is situated at the intersection of computer vision, generative modeling, and representation learning. She is recognized for her contributions to image and video generation, representation learning and active sensing. Her most recent work pushes the frontiers of vision and multi-modal generative models by devising novel model sampling, search, and post-training techniques to unlock high-utility and physically plausible synthetic data from generative models. Beyond her technical contributions, Adriana is an active member of the AI community: she has co-organized numerous workshops at top-tier conferences, including the ReGenAI workshop at CVPR (2024-2025) and sessions on the Science and Engineering of Deep Learning and Graph Representation Learning at NeurIPS and ICLR. Adriana earned her Ph.D. from the University of Barcelona under the supervision of Dr. Carlo Gatta, followed by a postdoctoral fellowship at Mila with Prof. Yoshua Bengio, where she advanced deep learning for complex biomedical and graph-structured data.
