roxie

Decisions for @CVPR 2026 are out—congratulations to all authors. I’m excited to share a community step forward: the new CVPR Findings Track. Area Chairs recommended 1717 papers for potential inclusion, creating a principled pathway to recognize and share valuable work that may not be the best fit for the main program—while still enabling authors to publish and present through integrated Findings poster sessions. As our field scales, we need not only better models—but better community infrastructure. This effort is led collectively by the Findings organizing team—Bryan Plummer, Kevin Shih, @anand_bhattad, @jccaicedo, @Grigoris_c, @BoqingGo, @liuziwei7, and me. Huge thanks to the CVPR General Chairs, Program Chairs, and especially the Area Chairs for supporting this step forward. Looking forward to seeing many of you at CVPR 2026—across the main program, Findings, and workshops.

Please check our work if you are interested in building an efficient generative model with semantic control!!!

The post below (learnable lambda has to do with skip connections) reminded me of some ResNet and Transformer architecture-related lore that I thought would be fun to write up! Blogpost: lb.eyer.be/s/residuals.ht…

Thank you @anand_bhattad very much for sharing our work!!

still the most compelling Figure 1 i've ever seen - from "Visualizing the Loss Landscape of Neural Nets" (2017)

Residual Connections Harm Generative Feature Learning! [1/4] Although residual connections have become the "gold standard" in network design, our research shows that they can be detrimental to unsupervised generative learning. Residual connections continuously propagate early features to the later layers, preventing the model from learning more abstract representations at the bottleneck. To address this, we (joint led by @propitious1235) propose the Decayed Identity Shortcut, which gradually reduces the impact of identity shortcuts, encouraging more abstract representations at the bottleneck. In experiments with diffusion models and MAE, our design significantly improves both representation learning and generation quality, without introducing any additional learnable parameters. Our analysis also reveals that this performance improvement is strongly associated with the emergence of lower-rank representations. abs: arxiv.org/abs/2404.10947 A huge thanks to my amazing collaborators: Ruoxi Jiang, Will Gao, Rebecca Willett, and Michael Maire!

Embed and Emulate: Contrastive representations for simulation-based inference. arxiv.org/abs/2409.18402

Delighted to share that our Latent Intrinsics paper has been accepted as a spotlight at #NeurIPS2024 ! Congratulations to the team and a big thank you to the reviewers and AC for their unanimous positive feedback. Stay tuned for more on latent intrinsics: a concept with broad applications in computer vision!🤓

I'm incredibly proud to share NeuralGCM, our new AI and physics based approach to weather and climate modeling with state-of-the-art accuracy, published today in @Nature: nature.com/articles/s4158…

9 Distance Measures in data science with algorithms (v/@gp_pulipaka).

@Delta they did not accommodate passengers from Miami to Chicago.They landed in Minneapolis last night and had kids sleeping on the floor! Couldn’t give hotels to the passengers, put them on other airline flights or refund their pay so they can travel home via car!#unprofessional

@TonyTheLion2500 Only this way: medium.com/towards-data-s…

The new National Institute for Theory and Mathematics in Biology will bring together mathematicians & biologists, with support of this exciting new opportunity thanks to the @NSF & @SimonsFdn. new.nsf.gov/news/new-50-mi…