Centre for Innovation in Computing at Lassonde

Close mobile menu

Marcus A. Brubaker

Assistant Professor, EECS Department
Member, Centre for Vision Research
Member, IC@L
Core Member, VISTA
Faculty Affiliate, Vector Institute
Faculty Member, NEXT AI
Status-Only Assistant Professor, CS Department, University of Toronto

Website | Email

2021 – 2022 Research Highlights

Invertible Transformations for Generative Modelling

This research program we explored invertible transformations for constructing novel forms of generative models.  Invertible transformations allow for constructing highly expressive generative models which are capable of efficient sampling and exact density modeling.  Further, by virtue of representing the transformation of a complex data distribution into a simple, known distribution, they allow for effective incorporation of domain specific knowledge.  In the past year I published the first introduction and review article on the topic of invertible generative models known as Normalizing Flows.  Further, I have published several papers exploring novel applications and models build upon normalizing flows, including the first normalizing flow model shown to model high resolution data.  I also ran a tutorial at ECCV 2020 on the topic.

Research Highlights

  1. Normalizing Flows: An Introduction and Review of Current Methods. Kobyzev, I.; Prince, S. J.; and Brubaker, M. A. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI). 2020.
  2. Wavelet Flow: Fast Training of High-Resolution Normalizing Flows. Yu, J. J.; Derpanis, K.; and Brubaker, M. A. In Neural Information Processing Systems (NeurIPS), 2020.
  3. Modeling Continuous Stochastic Processes with Dynamic Normalizing Flows. Deng, R.; Chang, B.; Brubaker, M. A.; Mori, G.; and Lehrmann, A. In Neural Information Processing Systems (NeurIPS), 2020.
  4. Tails of Lipschitz Triangular Flows. Jaini, P.; Kobyzev, I.; Brubaker, M. A.; and Yu, Y. In Proceedings of the International Conference on Machine Learning (ICML), 2020.
  5. Probabilistic Character Motion Synthesis using a Hierarchical Deep Latent Variable Model. Ghorbani, S.; Wloka, C.; Etemad, A.; Brubaker, M. A.; and Troje, N. F. Proceedings of Symposium on Computer Animation (SCA) in Computer Graphics Forum, 39(8). 2020.