←
Sergey Pozdnyakov139 posts
Sergey Pozdnyakov avatar

Sergey Pozdnyakov

@spozdn

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

◷ Joined May 2023▧ 8 Photos and videos
Jean-Philip Piquemal avatar
⟳ Sergey Pozdnyakov retweeted
Jean-Philip Piquemal@jppiquem·17 Dec 2025

💫 As promised, we just released on GitHub the weights of the #FeNNixBio1 foundation machine learning model for drug design! 💫 Weights: github.com/FeNNol-tools/FeNN… FeNNol GPU code: github.com/FeNNol-tools/FeNN… The models are distributed under the open source ASL license (i. e. restricted to non-commercial academic research). You can also check the updated version of the preprint that includes a unified transformers architecture as well as the full computation of the Freesolv hydration free energies dataset etc... doi.org/10.26434/chemrxiv-20… Happy holidays and merry Christmas everyone! 🎅 🎄 Sorbonne Université / CNRS @qubit_pharma #machinelearning #moleculardynamics #drugdesign #compchem #GPU #biophysics

Junwu Chen avatar
⟳ Sergey Pozdnyakov retweeted
Junwu Chen@JunwuChen25·10 Nov 2025

Thrilled to share our new work MatInvent, a general and efficient reinforcement learning workflow that optimizes diffusion models for goal-directed crystal generation. Thanks to @JeffGuo__ , @efssh , @pschwllr , @SchwallerGroup , @NCCR_Catalysis . arxiv.org/abs/2511.03112

Adil Kabylda avatar
⟳ Sergey Pozdnyakov retweeted
Adil Kabylda@kabylda_·15 Oct 2025

Our new work, “QCell: Comprehensive Quantum-Mechanical Dataset Spanning Diverse Biomolecular Fragments,” is now out on arXiv! 🌱 1/7

Post attachment
Philippe Schwaller (he/him) avatar
⟳ Sergey Pozdnyakov retweeted
Philippe Schwaller (he/him)@pschwllr·14 Oct 2025

Exciting postdoc opportunity in the @SchwallerGroup at EPFL! We're hiring a postdoc to advance ML-driven synthesis planning after Zlatko Joncev’s successful exit to co-found B-12 (YC '25) 🚀 Work on: - LLMs for strategic synthesis planning - Chemical reasoning at scale - Building the next-gen framework for retrosynthesis Our recent preprint shows that LLMs can guide synthesis planning with natural-language strategies — combining AI reasoning with traditional chemical tools (arxiv.org/abs/2503.08537). Join us at the intersection of chemistry & AI. Up to 2 years. Based in Lausanne 🇨🇭 Apply: forms.fillout.com/t/nnxVE3Rc… #ChemTwitter #MachineLearning #SynthesisPlanning #PostdocPosition

Xuan-Vu Nguyen avatar
⟳ Sergey Pozdnyakov retweeted
Xuan-Vu Nguyen@XuanVuNguyen18·12 Sep 2025

You don’t like molecular dynamics? We get it. That’s why at this year’s LLM hackathon for Chemistry and Materials Science, we built not one, but ✨two✨ AI agents for molecular dynamics 👇

1:57
Sergey Pozdnyakov avatar
Sergey Pozdnyakov@spozdn·10 Sep 2025

We did it for Convolutional Neural Networks, and it works too. lmKAN-based CNNs are Pareto optimal on both the CIFAR-10 and ImageNet-1k datasets, achieving 1.6-2.1× reduction in inference FLOPs at matched accuracy.

Post attachment
Sergey Pozdnyakov avatar
Sergey Pozdnyakov@spozdn·10 Sep 2025

High-dimensional linear mappings are a primary component of most architectures, not just fully connected neural networks. So, other architectures can also be made lmKAN-based.

Sergey Pozdnyakov avatar
Sergey Pozdnyakov@spozdn·10 Sep 2025

It works! Our lookup multivariate Kolmogorov-Arnold Networks (lmKANs) are consistently inference FLOPs Pareto optimal. For general function approximation, the inference FLOPs reduction is up to 6.0×. For the methane dataset, the H100 wall-clock time reduction is up to 12.9×.

Post attachment
Sergey Pozdnyakov avatar
Sergey Pozdnyakov@spozdn·10 Sep 2025

Furthermore, we introduce a multivariate extension to KANs, which is built upon multivariate functions as building blocks. Our layers can have hundreds of times as many trainable parameters as a linear layer with the same shape. Yet the inference FLOPs are only twice as large.

Post attachment
Sergey Pozdnyakov avatar
Sergey Pozdnyakov@spozdn·10 Sep 2025

Kolmogorov-Arnold Networks are natural hosts for this idea as they express general high-dimensional mappings in terms of multiple optimizable 1D functions that can be splined. We implement CUDA kernels for this approach.

Sergey Pozdnyakov avatar
Sergey Pozdnyakov@spozdn·10 Sep 2025

Spline lookup tables allow us to do better than that. Let's take an optimizable piecewise linear function defined on a regular grid with G intervals. It contains G + 1 independent trainable parameters. While G can be arbitrarily large, the computational cost to evaluate it at any given point is O(1), independent of the number of trainable parameters - one needs to compute just one linear term.

Sergey Pozdnyakov avatar
Sergey Pozdnyakov@spozdn·10 Sep 2025

The number of trainable parameters is a primary predictor of a model's capabilities. For a linear layer with dimensions N_in and N_out, the number of parameters is ~N_in N_out, and so is the number of multiplications. One trainable parameter = one multiplication at inference.

Sergey Pozdnyakov avatar
Sergey Pozdnyakov@spozdn·10 Sep 2025

You can have a fully connected layer with hundreds of times as many parameters as a linear layer of the same shape, but only 2× the inference FLOPs. CUDA kernels included. Paper: arxiv.org/abs/2509.07103 Code: github.com/schwallergroup/lm… #lmKAN Main idea below.

Sergey Pozdnyakov avatar
Sergey Pozdnyakov@spozdn·10 Sep 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!🧵