Research

Research and publications

My research is about making graph neural networks deeper and robots better at seeing and acting. Both started at TU Darmstadt.

Publication

NeurIPS 2024

NeurIPS 2024, Main Conference Track

Graph Neural Networks Need Cluster-Normalize-Activate Modules (opens in a new tab)

Arseny Skryagin, Felix Divo, Mohammad Amin Ali, Devendra Singh Dhami, Kristian Kersting

The paper introduces a plug and play module that clusters node features, normalizes each cluster and applies a learnable activation per cluster. It reduces oversmoothing and lets graph neural networks go deeper.

I am the third of five authors. The central idea, the method, the experiments and the results came out of my Master's thesis.

The paper reports 94.18 percent accuracy on Cora and 95.75 percent on CiteSeer, and lower error in regression tasks with fewer learnable parameters than competing architectures.

Cite this paper
@inproceedings{skryagin2024cna,
  title     = {Graph Neural Networks Need Cluster-Normalize-Activate Modules},
  author    = {Skryagin, Arseny and Divo, Felix and Ali, Mohammad Amin and Dhami, Devendra Singh and Kersting, Kristian},
  booktitle = {Advances in Neural Information Processing Systems},
  volume    = {37},
  year      = {2024},
  doi       = {10.52202/079017-2535}
}

Master's thesis

Selective Activations for Addressing Oversmoothing in Deep Graph Neural Networks

AI and Machine Learning Lab, TU Darmstadt. Supervised by Prof. Dr. Kristian Kersting. Submitted February 2024, graded 1.0.

The problem
When a graph neural network gets deeper, the features of its nodes become more and more alike. This is called oversmoothing, and it limits how deep these networks can usefully go.
What I did
I found that learnable rational activations alone were not enough. My method clusters the node embeddings, normalizes each cluster and then applies a learnable rational activation per cluster.
A new metric
The common MAD metric does not capture how spread out clusters are. I proposed MADC to measure cluster dispersion.
Experiments
Implemented in PyTorch and tested on Cora, CiteSeer, PubMed, DBLP and ogbn-arxiv.

The thesis received the 2025 prize for outstanding scientific achievement from the Vereinigung von Freunden der TU Darmstadt. Award page (opens in a new tab)

Robotics

Machine learning for robots

Jan 2020 to Mar 2023

Research Assistant, Machine Learning and Deep Learning for Robotics

Intelligent Autonomous Systems Lab (IAS), TU Darmstadt

  • Developed deep learning models for mobile robots, mainly for visual perception and obstacle avoidance with encoder decoder architectures.
  • Built model based reinforcement learning methods for adaptive robot behaviour and trained models for object recognition and scene understanding.
  • Developed robot software in ROS and trained and tested it in Gazebo and Isaac Gym, including grasp detection and motion planning with OMPL.
  • Prepared and supervised robot learning exercises for students in robotics, optimal control, machine learning and reinforcement learning.

Bachelor's thesis

An Educational Framework for Robot Learning

Intelligent Autonomous Systems Lab, TU Darmstadt. Supervised by Prof. Dr. Jan Peters. Submitted September 2021, graded 1.0.

I designed and built a Python robot simulator with exercises for students, using Pymunk, Pyglet and Sphinx. The exercises cover Denavit-Hartenberg, forward and inverse kinematics, splines, PID control, Markov decision processes with policy and value iteration, ProMPs, REPS and ball catching.