Piotr Tempczyk

About me

I am an AI researcher and lead the Sigma research group at IDEAS Research Institute in Warsaw. We study machine learning for complex adaptive systems, including financial markets and energy, with a focus on non-stationarity, low signal-to-noise ratios and emergent behaviour. I hold a PhD in Computer Science from the University of Warsaw, where I conducted my doctoral research under Marek Cygan's supervision. My thesis explored how to estimate the local intrinsic dimension of data — its effective number of degrees of freedom — using neural density models. My academic work includes publications at ICML, AAAI and ICLR.

I have worked in machine learning since 2012. My commercial experience includes serving as Head of AI Research at TradeLink (now DV Trading), where I combined hands-on deep learning research for financial markets with leadership of an AI research team. Earlier roles include Co-Founder & CTO of Deeptale.ai and Head of AI at AI Clearing, SonarHome and Nethone (within Daftcode), spanning generative models, computer vision, property valuation and fraud detection.

I welcome research collaborations, speaking invitations and selected consulting and advisory work with companies and startups. I can help assess AI technologies, shape research plans and advise on building AI teams. Get in touch on LinkedIn.

Research interests

  • Deep learning for financial markets
  • Machine learning for complex adaptive systems
  • Local intrinsic dimension, data geometry and topology
  • Generative models and density estimation
  • Bayesian statistics, causal inference and probabilistic machine learning

Publications

Selected publications and research projects, including my doctoral research and collaborations:

ICLR 2026 poster presenting benchmarks and failure modes of local intrinsic dimension estimators.

Why We Need New Benchmarks for Local Intrinsic Dimension Estimation

ICLR 2026
Piotr Tempczyk, Dominik Filipiak, Łukasz Garncarek, Ksawery Smoczyński and Adam Kurpisz.

We introduce benchmarks for local intrinsic dimension estimators, using controlled transformations that preserve manifold structure and expose failure modes. Success on simple synthetic datasets does not necessarily transfer to realistic domains.

Paper · Code · Poster

LIDL: Local Intrinsic Dimension Estimation Using Approximate Likelihood

Our paper with R. Michaluk, Ł. Garncarek, P. Spurek, J. Tabor and A. Goliński on local intrinsic dimension estimation was selected for an oral presentation at ICML 2022.

NeurIPS 2021 BEETL Competition: Benchmarks for EEG Transfer Learning

We took 3-rd place in the competition with Maciej Śliwowski and Vincent Rouanne. Our work was presented during the talk at the NeurIPS 2021 BEETL competition workshop.

n-CPS semi-supervised segmentation research

n-CPS: Generalising Cross Pseudo Supervision to n Networks

IEEE Access, 2026
Dominik Filipiak, Piotr Tempczyk and Marek Cygan.

We extend cross pseudo supervision to multiple networks for semi-supervised semantic segmentation, combining predictions through ensembling. Originally released as a preprint in 2021.

Paper · Original preprint (2021)

One Simple Trick to Fix Your Bayesian Neural Network

Preprint, 2022

Joint work with K. Smoczyński, P. Smolenski-Jensen and M. Cygan on how different activation functions in Bayesian neural networks affect posterior shape, network calibration and uncertainty estimates. The preprint is available on arXiv.

Reinforcement learning based brain-computer interface

Joint work with Maciej Śliwowski. Poster presentation at EEML 2020 and best poster award in reinforcement learning category.

Teaching

These are the events and courses I was involved in:

Contact Me

Email: piotr.tempczyk at ideas.edu.pl