Vector Researcher Will Grathwohl Aims to Lower AI Entry Barriers with FFJORD
Will Grathwohl, a researcher at the Vector Institute and graduate student at the University of Toronto, is working to make artificial intelligence more accessible by reducing its high costs and data requirements. Drawing parallels to the early internet, Grathwohl argues that AI's potential will only be realized when barriers are lowered for non-experts. At the International Conference on Learning Representations (ICLR), he presented FFJORD, a paper co-authored with colleagues including David Duvenaud. FFJORD utilizes continuous time dynamics to improve normalizing flows, a type of generative model. This approach allows for better use of unlabeled data, addressing the industry's reliance on expensive, hand-labeled datasets and massive computing power. The technology builds upon previous award-winning research on Neural Ordinary Differential Equations. Grathwohl, inspired by his industry experience where data labeling was a significant bottleneck, believes that efficient unsupervised generative models are key to democratizing AI. This advancement could eventually aid fields like genetics and robotics by enabling complex modeling with fewer resources, ultimately putting powerful AI tools into the hands of those with innovative application ideas rather than just technical creators.
Wire timeline
Vector Researcher Will Grathwohl Aims to Lower AI Entry Barriers with FFJORD
Will Grathwohl, a researcher at the Vector Institute and graduate student at the University of Toronto, is working to make artificial intelligence more accessible by reducing its high costs and data requirements. Drawing parallels to the early internet, Grathwohl argues that AI's potential will only be realized when barriers are lowered for non-experts. At the International Conference on Learning Representations (ICLR), he presented FFJORD, a paper co-authored with colleagues including David Duvenaud. FFJORD utilizes continuous time dynamics to improve normalizing flows, a type of generative model. This approach allows for better use of unlabeled data, addressing the industry's reliance on expensive, hand-labeled datasets and massive computing power. The technology builds upon previous award-winning research on Neural Ordinary Differential Equations. Grathwohl, inspired by his industry experience where data labeling was a significant bottleneck, believes that efficient unsupervised generative models are key to democratizing AI. This advancement could eventually aid fields like genetics and robotics by enabling complex modeling with fewer resources, ultimately putting powerful AI tools into the hands of those with innovative application ideas rather than just technical creators.
Vector Institute for Artificial Intelligence