Hi! I'm a machine learning scientist with a software engineering & mathematics background.
Currently working at Plumerai on efficient & private AI. Previously, my PhD research focused on hyperparameter optimization for deep learning via evolutionary algorithms.
I'm excited to apply my expertise to problems that matter, using whichever tool fits best: classic ML, deep learning, or generative AI. I find evaluation and safety work ever more important.
My hobbies include reading (non-fiction / fantasy / diverse), hiking, poetic translation, and discussing weird ideas. Happy to chat with strangers, so get in touch!
-
Iterated Population Based Training with Task-Agnostic Restarts [ICML]
TL;DR We introduce a PBT variant that uses task-agnostic restarts to automatically adjust its step size and efficiently optimize hyperparameter schedules. | Paper | Code
- Subscribe to the technical blog via RSS - I recommend CommaFeed
-
To Be Greedy, or Not to Be - That Is the Question for Population Based Training Variants [TMLR]
TL;DR Bayesian PBTs optimize the greedy objective more effectively than non-Bayesian PBTs, this can be good or bad (depends on the task & hyperparams) | Paper | Code
-
Transformer explanations: a collection
Transformer is a powerful architecture that can be difficult to understand. There are many great explanations on the web, each approaching the subject in a different way. Here I link the explanations I liked, and mention who I believe the target audience is for each one.
-
PBT-NAS: mix architectures during training
Based on our paper “Shrink-Perturb Improves Architecture Mixing during Population Based Training for Neural Architecture Search” (ECAI 2023)
-
Install nvcc without root
I needed to install
nvccon our group server, where I lack root privileges. I found a nice guide, in this post I will slightly expand on it by explicitly mentioning every step I had to take. -
ENCAS: Search cascades of neural networks in any model pool
Based on our paper “Evolutionary Neural Cascade Search across Supernetworks” (Best Paper award @ GECCO 2022)
TL;DR Search for cascades of neural networks in model pools of hundreds of models, optionally using Neural Architecture Search to efficiently generate many diverse task-specific neural networks. Get dominating trade-off fronts on ImageNet and CIFAR-10/100, improved ImageNet SOTA (of publicly available models).
-
Document recognition with Python, OpenCV, and Tesseract
Recently I’ve conducted my own little experiment with the document recognition technology: I’ve successfully went from an image to the recognized editable text.
-
Not-quite-ivory towers
- Subscribe to the general blog via RSS - I recommend CommaFeed
-
Translation trade-offs
- 2026-07 Дом Ветров Отражение → 🇬🇧 Reflection
- 2026-01 Robert W. Service The Cremation of Sam McGee 🇬🇧 → Кремация Сэма МакГи
- 2025-06 Ноль Человек и кошка → 🇬🇧 A Kitty and a Person
- 2024-04 De 4Tuoze Matroze Wie aan Amsterdam komt 🇳🇱 → Кто Амстердам хоть тронет
- 2023-12 ★ Heidevolk Yngwaz' Zonen 🇳🇱 → Сыны Ингвы
- 2023-11 ★ Юрий Энтин До чего дошёл прогресс → 🇬🇧 Look at All That Progress Bore
- 2023-08 Miracle of Sound Into the Wild 🇬🇧 → В дикий простор
- 2023-04 Машина времени Однажды мир прогнётся под нас → 🇬🇧 One Day the World Will Bow Before Us
- 2023-02 Paul and Storm Write Like the Wind (George R. R. Martin) 🇬🇧 → Как ветер пиши (Джордж Р Р Мартин)
- 2022-12 Zupfgeigenhansel Ich bin Soldat 🇩🇪 → Пусть я солдат
- 2022-01 Miracle of Sound Zombie Holiday 🇬🇧 → Отпуск средь зомбей
- 2017-01 Tall like Tyrion The Dornishman's Wife 🇬🇧 → Жена дорнийца
- Cities where I lived at least a month: St. Petersburg, Cologne, Berlin, Amsterdam, Leiden
- Favorite color: Orange
- What DEFINITELY DID NOT influence my decision to move to 🇳🇱: my favorite color
- Most steps walked in a day: 35,228
- An octopus has more neurons in its tentacles than in its brain 🐙