Sergey Pozdnyakov

@spozdn

Postdoc @SchwallerGroup @EFFL | Geometric deep learning on 3D point clouds for atomistic modeling

Lausanne Joined May 2023

517 Following 342 Followers 139 Posts

Latest posts 5

Sergey Pozdnyakov @spozdn Oct 19, 2025

See you there! 🤗

Hannes Stark @HannesStaerk Oct 19, 2025

Reading group session on Monday: "Lookup multivariate Kolmogorov-Arnold Networks" arxiv.org/abs/2509.07103 with Sergey Pozdnyakov

On zoom at 9am PT / 12pm ET / 6pm CE(S)T: portal.valencelabs.com/starklyspeaking

View quoted post on X

Sergey Pozdnyakov @spozdn Sep 10, 2025

High-dimensional linear mappings, or linear layers, dominate both the parameter count and inference cost in most deep learning models. We propose a general-purpose drop-in replacement with a substantially better capacity - inference cost ratio. Check it out!🧵

Sergey Pozdnyakov @spozdn Jan 16, 2025

Sergey Pozdnyakov @spozdn Nov 27, 2024

It was a lovely symposium! Enjoyed it much

Miroslava Nedyalkova @MiroslavaNedya1 Nov 26, 2024

@spozdn Best Poster @unifrChemistry Great work!!

View quoted post on X

Sergey Pozdnyakov @spozdn Sep 5, 2024

Remarkable empirical evidence of the efficiency of unconstrained models!

Mark Neumann @MarkNeumannnn Sep 3, 2024

We are using a simple architecture, augmented with attention, so these models scale very well (5x faster than MACE with a system of 10k atoms).

In particular - our models are NOT equivariant. Our models hopefully show that this is a modeling choice, not a requirement for Matsci.
View quoted post on X