Alexandre Galashov
@agalashov
Followers
511
Following
748
Statuses
64
Machine Learning Research @ Deepmind
London, England
Joined April 2013
RT @Ar_Douillard: We release today the next step for distributed training: --> Streaming DiLoCo with Overlapping Communication. TL;DR: tr…
0
87
0
Diffusion models are trained to predict the mean E[X_0 | X_t] of clean data given a noisy sample. We propose a novel method of learning the full posterior distribution p(X_0 | X_t) using scoring rules, which improves performance of few-step diffusion models. Check our paper - Joint work with @ValentinDeBort1 Guntupalli @zhouguangyao @sirbayes @ArthurGretton
@ArnaudDoucet1
Better diffusions with scoring rules! Fewer, larger denoising steps using distributional losses; learn the posterior distribution of clean samples given the noisy versions. @agalashov @ValentinDeBort1 Guntupalli @zhouguangyao @sirbayes @ArnaudDoucet1
0
20
119
RT @ArnaudDoucet1: A standard ML approach for param estimation in latent variable models is to maximize the expectation of the log of an im…
0
26
0
Check out our new work on speculative sampling for diffusion models
Speculative sampling accelerates inference in LLMs by drafting future tokens which are verified in parallel. With @ValentinDeBort1 , @agalashov & @ArthurGretton, we extend this approach to (continuous-space) diffusion models:
0
1
12
RT @ArnaudDoucet1: Speculative sampling accelerates inference in LLMs by drafting future tokens which are verified in parallel. With @Valen…
0
50
0
0
0
1
Check out this amazing work by @RishabhKabra !
VLMs do *not* describe 3D objects consistently from different angles! But we could summarize different descriptions using a language model, right? Well… here’s what happens when an LLM merges multiple contradictory descriptions of the same object.
0
0
1
Great opportunity at Google DeepMind in an excellent team!
My team is looking for a research engineer in New York! Our recent efforts include DiLoCo (distributed learning) and DiPaCo (distributed mixture of experts). Those projects that I've co-led, were the most exciting projects I've contributed, and i can tell you one thing: there is more to be done.
0
3
12
Check out our new pre-print on using forward diffusion to train MMD GAN discriminator without adversarial training. To sample -- just use noise-adaptive MMD gradient flow!
Do you have a GAN critic? Then you have a diffusion! Divergence is✨adaptive✨ MMD GAN critic. No adversarial training needed: just add noise to your data sample as in a classic diffusion. Figure: adaptive (ours) vs fixed. @agalashov @ValentinDeBort1
0
4
23
0
0
1
1
0
2
RT @Ar_Douillard: I’m honored to see our work on distributed training (DiLoCo) and distributed mixture of experts (DiPaCo) highlighted duri…
0
1
0
Work done together with Michalis Titsias , @AmalRannen , Razvan Pascanu, @yeewhye , Jorg Bornschein !
0
0
1
Work done with Jorg Bornschein, Ross Hemsley, @amalrannen , @yutianc , @arslan_mac , Owen He, @ar_douillard , @MassCaccia , @qixuan_feng , @jiajun_s , @SRebuffi , @kittystacpoole , @diegolascasas , @willhawkins3 , @aggielaz , @yeewhye , @andreialexrusu , @rpascanu, @MarcRanzato
0
0
4