Hi there, I'm Bogdan Ivanyuk-Skulskyi

Bogdan photo

I am a machine learning engineer at Cyclop.ai (Vinci), working on security systems for road infrastructure (tunnels, bridges, etc.) and autonomous vehicles.

I obtained my Bachelor's degree in 2020 and completed my Master's degree in 2022 at the National University of Kyiv-Mohyla Academy. Throughout my academic journey, I had the privilege of working under the guidance of Prof. Galyna Kriukova, Andrii Dmytryshyn, and Nadiya Shvai.

During my Master's program, I had an internship at Samsung Research, specifically within the Integent Security Lab, focusing on voice security applications. Additionally, I did an internship at the University of Toronto, where I worked under the supervision of Dr. Brokoslaw Laschowski, and Dr. Alex Mihailidis on video classification model for exoskeleton control.

My interests primarily include video processing, anomaly and out-of-distribution data detection, reinforcement learning, and robotics.


Publications:

2024:
  • Towards Lightweight Transformer Architecture: an Analysis on Semantic Segmentation
    B Ivaniuk-Skulskyi, N Shvai, A Llanza, A Nakib
    In ACDSA 2024, with oral presentation
    [paper]
    Examining modifications to the Transformer model for semantic segmentation, two approaches were explored: skip-attention and pool-unpool attention. Skip-attention improved inference speed by +26.7% for ADE20k and +101.7% for Cityscapes, while pool-unpool attention showed gains of +14.8% for ADE20k and +73.3% for Cityscapes.

2023:
  • StairNet: Visual recognition of stairs for human-robot locomotion
    A G Kurbis, D Kuzmenko, B Ivaniuk-Skulskyi, A Mihailidis, B Laschowski
    BioMedical Engineering OnLine
    [paper]
    Comparison of state estimation models of the walking environment in robotic prosthetic legs and exoskeletons.

  • Sequential Image Classification of Human-Robot Walking Environments using Temporal Neural Networks
    B Ivaniuk-Skulskyi, A G Kurbis, A Mihailidis, B Laschowski
    In IEEE ICRA 2023, Computer vision for Wearable Robotics workshop; KITE/UHN ICAIR 2023
    [ICAIR Poster] [ICRA Abstract] [pre-print] [Code]
    State estimation model of the walking environment in robotic prosthetic legs and exoskeletons, focusing on dynamic human-robot walking dynamics rather than static image classification.
2020:
  • Geometric Properties of Adversarial Images
    B Ivaniuk-Skulskyi, G Kriukova, A Dmytryshyn
    In IEEE DSMP 2020
    [Paper] [Code]
    Linear algebra-based approach to detect adversarial images, addressing the challenge of identifying slight modifications that mislead machine learning models' predictions, and supporting the method with theoretical explanation and numerical experiments.
    Megaverse gif


Selected open-source projects:

  • Tower Defence RL agent [Code]
    An implementation of a classical Tower Defence game, which allows you to play it on your own, and it also includes a gym-like environment for training RL agents.
    Surviv.io rl agent
  • Raspberry Pi car [Code]
    Building a Raspberry Pi car, from putting it together to testing and adding remote control features.
    GIF 1 GIF 2
  • UA-datasets [Documentation] [Code] [Hugging Face Hub]
    Collection of datasets for the Ukrainian language. These datasets include text classification, question answering, and token classification tasks. They are neatly organized and provided within a Python package and also accessible on the Hugging Face Hub for easy use and integration into your NLP projects.
  • Surviv.io game bot [Code]
    Reinforcement Learning Agent for the Online Multiplayer Battle Royale Game surviv.io.
    Surviv.io rl agent
  • Quora Insincere Questions Classification (Kaggle) [Code]
    NLP competition for insincere questions classification (silver medal, top 2% solution).
  • Exploring cross-lingual abilities of Multilingual BERT [Code]
    I studied M-BERT's transfer learning for classification tasks, moving from English to Russian test sets, and compared the results to RuBERT. I also explored M-BERT's transfer learning for Boolean Questions, fine-tuning on English BoolQ and NLI datasets, then validating on a Russian test set and comparing the outcomes to RuBERT.

Let's get in touch!