We will be presenting at the poster session next week on Dec 13th, 5-7 pm CT. You can also check out our virtual poster at: https://t.co/1bdjSaX4JO. Thanks for your attention :)

Our design allows one to vary the metric geometry of a given manifold neural network and directly note the effect while fixing network architecture. We use this in practice to test SPD manifold networks with different Riemannian metrics (Affine-Invariant and Log-Euclidean)!

We propose a novel, principled, and geometrically natural generalization of residual neural networks to general Riemannian manifolds. The construction relies only on knowledge of geodesics as specified by a manifold's exponential map, a natural construct for residual addition.

I am elated to present our new NeurIPS 2023 paper, "Riemannian Residual Neural Networks," together with @ericmchen1, @sholalkere, Anna Asch, @aaron_lou, @sernamlim, @chrismdesa! Arxiv: https://t.co/ovJ5R0wJPo Github: https://t.co/gfHMxXQzm2

If you want to meet up, feel free to DM me on Twitter! Looking forward to reconnecting :)

Great to be with friends at the first in-person conference in 3 years! @YifeiZhou02 @sidhanth_h @ericmchen1 @aaron_lou

Happy to announce that I will be joining @Yale to do a PhD in applied mathematics with Prof. Anna Gilbert @annacgilbert after I graduate from @Cornell! A big thank you to my collaborators, advisors, friends, and family for helping me throughout this journey.

@aaron_lou @dereklim_lzh @sernamlim @chrismdesa We will be presenting Equivariant Manifold Flows at NeurIPS 2021 shortly, at the 7:30-9:00 EST poster session. Come and chat with us! :) We're at Poster Spot E2 in the gather town. Paper: https://t.co/S1GkKikgcz Code: https://t.co/Wl0B9Mnhh6

@aaron_lou @dereklim_lzh @sernamlim @chrismdesa Our work fits into the broader scope of normalizing flows on manifolds. This was a fun project to work on; if you’re interested in learning more, please join us at ICML INNF 2021 this Friday, July 23rd at 7:00-8:00 EST!

@aaron_lou @dereklim_lzh @sernamlim @chrismdesa Our construction is general and outperforms prior work due to our continuous flow structure. ​​We parameterize invariant potentials by neural networks, and use the induced equivariant gradient fields in the neural manifold ODE frameworks of NMODE and RCNF to learn flows.

@aaron_lou @dereklim_lzh @sernamlim @chrismdesa Inspired by SU(n) gauge equivariant flows ( https://t.co/p5fF45xrxT) and equivariant flows in Euclidean space ( https://t.co/durvjgVUkC), we extend equivariant flows to arbitrary smooth manifolds where symmetry is with respect to action by any isometry subgroup.

I am happy to present our new work, “Equivariant Manifold Flows”, together with @aaron_lou, @dereklim_lzh, Qingxuan Jiang, @sernamlim, @chrismdesa! Arxiv: https://t.co/S1GkKikgcz

Congratulations to the Cornell University AI undergraduate club for 2 papers accepted at #NeurIPS2020! And thank you to @FacebookAI for your continued support of @CUAI_Cornell

Two papers accepted at Neurips!

We'll be presenting Neural Manifold ODEs at the INNF workshop at #ICML2020 at 1pm EST. https://t.co/YBvVwUBYDA . Come if you want to see how to learn complex densities on manifolds!

We just released the code for our #ICML2020 paper "Differentiating through the Fréchet Mean"! Check it out at https://t.co/BJcvLPpTaa. Also, come chat with @aaron_lou and me at 9 PM EST today at https://t.co/E7zm0tzhGn.

Here's another high quality gif of a learned density on H^2. This is `bigcheckerboard`; the above was `5gaussians` learned on H^2. To reproduce these results and see the groundtruth densities, check out our repo at https://t.co/xk9UbHcv0I.

Interestingly, we found our method was much more sample efficient than existing discrete normalizing flow methods, and was able to learn the overarching structure of fairly complex densities early on (decent even at ~60,000 samples out of 5,000,000 used for the final result).

We just released a paper on the manifold extension of Neural ODEs, enabling manifold continuous normalizing flows (tractable densities on manifolds are needed in physics and robotics, e.g. for unitary Lie groups in lattice QFT). Check out our paper ( https://t.co/PB1SR4PFmQ)!

If you work in geometric deep learning, you may find some of our methods interesting :) come to our session at (virtual) ICML and talk with us!