Théo Uscidda
@theo_uscidda
Followers
255
Following
511
Media
5
Statuses
58
PhD @ENSAEparis | past @AmazonScience, @FlatironInst, @fabian_theis lab @HelmholtzMunich.
New York, NY
Joined October 2023
Curious about the potential of optimal transport (OT) in representation learning? Join @CuturiMarco's talk at the UniReps workshop today at 2:30 PM! Marco will notably discuss our latest paper on using OT to learn disentangled representations. Details below ⬇️
2
28
180
RT @zeynepakata: It is a great honor to receive the ZukunftsWissen Prize 2025 from the German Academy of the Sciences @Leopoldina with gene….
0
6
0
RT @xavierjgonzalez: @NeurIPSConf overleaf has crashed, any chance we could just merge the full paper and supplemental deadlines, for a sin….
0
2
0
RT @LucaEyring: Catch me at #ICLR2025 today - I’ll be presenting our work on Quadratic OT for Representation Learning, the Gromov-Monge Gap….
0
3
0
RT @ExplainableML: @mwbini @shuchen_wu (4/4) Disentangled Representation Learning with the Gromov-Monge Gap.@LucaEyring will present GMG,….
0
2
0
RT @fabian_theis: 1/ Excited to share CellFlow, a new approach for complex perturbation modeling in single-cell genomics based on flow matc….
0
70
0
RT @ExplainableML: (3/4) Disentangled Representation Learning with the Gromov-Monge Gap.A fantastic work contributed by @theo_uscidda and @….
0
2
0
Our work on geometric disentangled representation learning has been accepted to ICLR 2025! 🎊See you in Singapore if you want to understand this gif better :).
Curious about the potential of optimal transport (OT) in representation learning? Join @CuturiMarco's talk at the UniReps workshop today at 2:30 PM! Marco will notably discuss our latest paper on using OT to learn disentangled representations. Details below ⬇️
0
18
154
RT @ArnaudDoucet1: Speculative sampling accelerates inference in LLMs by drafting future tokens which are verified in parallel. With @Valen….
0
51
0
RT @CNRSinformatics: #Optimisation | Gabriel Peyré, directeur de recherche CNRS au DMA, intervient lors de la conférence optimisation pour….
0
6
0
RT @mathSTb: Nayel Bettache: Bivariate Matrix-valued Linear Regression (BMLR): Finite-sample performance under Identifiability and Sparsity….
0
4
0
The best resource I’ve found so far on unifying flow matching and diffusion!.
A common question nowadays: Which is better, diffusion or flow matching? 🤔. Our answer: They’re two sides of the same coin. We wrote a blog post to show how diffusion models and Gaussian flow matching are equivalent. That’s great: It means you can use them interchangeably.
0
2
16
This is a joint work with the amazing.@LucaEyring @confusezius @fabian_theis @zeynepakata @CuturiMarco. Check out the paper!
0
3
9
We demonstrate the effectiveness of this approach across standard benchmarks. Additionally, we provide both empirical and theoretical evidence that preserving angles leads to better disentanglement compared to preserving (scaled) distances.
1
1
6
To get the best of both worlds, we rely on Gromov-Monge maps, trading mappings that *fully* preserve geometric features for ones that preserve them *as much as possible*. To achieve this, we introduce the Gromov-Monge Gap, a regularizer that transforms any NN into such a map.
1
2
7
To get the best of both worlds, we rely on Gromov-Monge maps, trading mappings that *fully* preserve geometric features for ones that preserve them *as much as possible*. To achieve this, we introduce the Gromov-Monge Gap, a regularizer that transforms any NN into such a map.
0
0
1
However, matching the prior while preserving geometric features remains challenging, as a mapping that fully preserves these features while aligning the data distribution with the prior typically does not exist….
1
1
4
Disentanglement is often achieved through prior matching. Moreover, it can be enhanced by structuring the latent space to preserve geometric features of the data, such as distances or angles between data points.
2
2
9
RT @RednasTom: 🎉Exciting news from @AIatMeta FAIR! We've released a Watermark Anything Model under the MIT license!.It was announced yester….
0
4
0