🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

The link between diffusion models and optimal transport is still a bit of an enigma to me. One thing that's clear: different diffusion models trained on similar datasets tend to recover similar mappings. If these are generally not OT, in what sense are they optimal instead?

✍️This paper is a really interesting read: arxiv.org/abs/2505.19712! We know diffusion models aren't optimal transport in general, but what if we use a linear interpolation schedule (e.g., flow matching one) and apply Reflow multiple times? Will it eventually converge to an OT

New video in the Generative Memory Lab channel! Krunoslav Lehman Pavasovic (Meta & ENS Paris) presented his work: "Classifier-Free Guidance: From High-Dimensional Analysis to Generalized Guidance Forms" youtu.be/94mXzub4JRc?si…

Zeros of random polynomials with i.i.d. coefficients exhibit interesting convergence properties. In particular, normal coefficient leads to weak convergence of the empirical measure to the unit circle.

🚨🚀 Come hear Grigory Bartosh talk about SDE Matching! SDE Matching is a highly efficient and scalable training framework for Latent/Neural SDEs. You no longer have to discretize or simulate your SDE models when fitting them to data. #SDE #Diffusion #FlowMatching #ML

📢📢 Elucidated Rolling Diffusion Models (ERDM) How can we stably roll out diffusion models for sequence generation in data-scarce dynamical systems? We elucidate the design of rolling diffusion, inspired by prob. flow ODEs and nonisotropic noise. 📄 arxiv.org/pdf/2506.20024

Why does RL struggle with tasks requiring long reasoning chains? Because “bumping into” a correct solution becomes exponentially less likely as the number of reasoning steps grows. We propose an adaptive backtracking algorithm: AdaBack. 1/n

Oh no language models are just memorization machines. They can’t possibly recover hidden rules! They are just series of matrix multiplications! Matrix multiplications can’t be used to efficiently represent complex algebraic structures, right?

Memory is a fundamental aspect of human cognition, yet current state-of-the-art AI models use only its rudimentary forms. Join us at the ICCV 2025 workshop on Memory & Vision in sunny Honolulu, where we will explore the intersection of memory and visual AI. We invite you to

Here is my short and practical thread on how to teach your Text-to-Image model to generate readable text: [1/n]

We desperately need to take all the people in SF who talk about accelerating biology but have only ever done math or CS and have them do a six month rotation in a wet lab.

Hyperbolic surfaces are impossible to imagine, but aspects of their geometry are still mathematically accessible. To study these mysterious shapes, mathematicians measure how connected they are. Here’s what that means: quantamagazine.org/years-after-th…

The popularity of this kind of ignorant posts is the direct result of Youtubers irresponsibly pushing the "scientists make shit up for money" line.

It has been a pleasure to be interviewed for this beautiful article on generalization in diffusion models I do think that the work is an important step in the right direction

[1/5] Data often has inherent geometric and topological properties that can be exploited by machine learning algorithms. Below are four areas of machine learning that explore this important aspect.

For decades, leading universities have operated in a recruiting echo chamber, hiring faculty near exclusively from each other. The tech industry is now creating its own echo chamber for AI talent. In both cases, this inbreeding suffocates innovation and impedes genuine progress

Yeah, Ill-posed problems are racist

Excited to share our #ICML2025 Spotlight paper "On the Guidance of Flow Matching", a framework of exact energy guidance in flow matching models. This work theoretically studies how to achieve controlled generation in the powerful generative model, flow matching, and proposes an