Diffusion models turn the data into a mixture of isotropic Gaussians, and so struggle to capture the underlying structure when trained on small datasets. In our new #ECCV2024 paper, we introduce RS-IMLE, a generative model that gets around this issue. Website:

More results are below. This was a fun collaboration with @_Linjunru, @researchirag, @mikacuy, @Stearns2Colton, @XuanLuo14 and @GuibasLeonidas :) 5/5

As shown, the approach significantly outperforms recent baselines. 4/5

Our approach, Global Motion Corresponder (GMC), learns to match points globally and is designed to maximize consistency between local point configurations across the different states. 3/5

Scenes with large motion present a major challenge, since the corresponding point may not be inside a local window. Existing methods typically assume motion to be small and fail under large motion. 2/5

If you are at #ICCV2025, check out our work on interpolating between two states of a 3D scene with large motion at poster 269 on Tuesday afternoon. It proposes a general-purpose method that can disambiguate points with similar appearance. Website: https://t.co/Kcx0JFEL3U 1/5

How can LLMs be made to handle longer contexts efficiently? Most prior methods require retraining - can we do without? In IceFormer, we showed how by repurposing Prioritized DCI, a nearest neighbour search algorithm. Find out more at @Mao_Yuzhen’s oral at the ICML LCFM workshop

What a day at the Vision & Learning Workshop at #ICML2025! With an incredible lineup of speakers to lighting talks on recent advancements in machine learning, researchers shared insights into the future of AI research. A huge thank you to everyone who made it a success!

Interested in how to generate realistic human trajectories using diffusion with just one step🤔? This is now possible with MoFlow, a one-step Flow Matching method accompanied with Implicit Maximum Likelihood Estimation based distillation. 🚀 Join us at #CVPR2025 in Nashville!

Are Diffusion and Flow Matching the best generative modelling algorithms for behaviour cloning in robotics? ✅Multimodality ❌Fast, Single-Step Inference ❌Sample Efficient 💡 We introduce IMLE Policy, a novel behaviour cloning approach that can satisfy all the above. 🧵👇

This looks like a game-changer for robotics and imitation learning! Diffusion policy is hard to run in real-time; the paper proposes IMLE policy, which can run at >110 Hz (a 61x speedup), use 38% less data and achieve better success rates. Note: I wasn't involved in this work,

It breaks my heart to learn of the tragic loss of a brilliant AI scientist and colleague. In honouring his memory, here's hoping that we as a community will engage in more open conversations around mental health and support one another. Many of us have come face to face with

Looking forward to #NeurIPS2024, happening right here in #Vancouver, BC, from Dec. 10-15. @SFU Computing Science is proud to have 7 papers at this year's conference! https://t.co/eiokJkVOP0

@geoffreyhinton @NobelPrize Here's the link for those who missed it:

Super inspiring example from @geoffreyhinton's @NobelPrize lecture - it shows how one could account for ambiguity in 3D scene understanding and model multiple plausible interpretations with Boltzmann machines. Interesting food for thought for computer vision researchers...

@SFU More details are in our job posting: https://t.co/oYd1VApfRZ

Interested in tackling hard problems? Passionate about helping spark the next leap in technology? Dreamt of coming to beautiful Vancouver? Now’s your chance! SFU CS is recruiting two assistant or associate professors in cybersecurity, NLP and/or software engineering, all super

Congrats to the #NobelPrize2024 winners for AI-driven breakthroughs in protein structure prediction! Learn how TTIC's Prof. Jinbo Xu's (@jinboxu_chicago) pioneering work in #deeplearning helped lay the foundation for these groundbreaking discoveries: https://t.co/1RS3vqu8Fn

My first exposure to generative models was through @geoffreyhinton’s course at @UofT, and my very first research project was on restricted Boltzmann machines as an undergrad in @zemelgroup. So thrilled and excited to see Hopfield networks and Boltzmann machines recognized with

RS-IMLE can be trained on as few as 100-200 images and improves the state-of-the-art by 35-62% in terms of FID. For more details, check out https://t.co/sEk922GyNr (6/6)