Dmitry Kobak

dmitry.kobak@uni-tuebingen.de

Google Scholar | Institute webpage

ORCID Github Pubmed Dblp

@hippopedoid on Twitter
@amoeba at CrossValidated.SE

I am a research scientist and a group leader in the Berens lab at Tübingen University, Germany, working on machine learning and data science for biological applications.

I am interested in self-supervised and unsupervised learning, in particular contrastive learning, manifold learning, and dimensionality reduction for 2D visualization of scientific datasets. I am working with image data, text data, graph data, and single-cell RNA-seq data in neuroscience contexts.

I am also interested in statistical forensics and have been involved in the analysis of Russian electoral falsifications, Russian war fatalities, and Covid-19 excess mortality.

In 2023/2024 winter semester I was a Vertretungsprofessor (visiting professor) at Heidelberg University, Germany. I am a member of the ELLIS society, a member of the Cluster of Excellence "Machine Learning for Science", and an IMPRS-IS associated scientist.

Teaching

Winter semester 2023/24 (Heidelberg): BSc course Einführung ins Machinelle Lernen (in German), MSc seminar Transformers, large language models, and their use in physics.

2020–2023 (Tübingen): For several years I was teaching an introductory course on machine learning for MSc students in neuroscience and data science. In winter semester 2020/21, due to the Covid pandemic, the class was held online and the lectures were recorded in a studio.

• Videos: Tübingen Machine Learning / Introduction to Machine Learning
• Slides: 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11

Talks

• Contrastive and neighbor embedding methods for data visualisation, April 2023, Network Seminar.
• Контрастные методы и методы ближайших соседей для визуализации данных, May 2023, SciBerloga Seminar.
• What are 2D neighbour embeddings of scRNA-seq data actually useful for?, April 2022, ICLR GTRL workshop.
• Neighbour embeddings for scientific visualization, June 2021, ELLIS Life Heidelberg.

Supervision / team

• Sebastian Damrich (started in 2023)

• Rita González Márquez (started in 2022)
• Niklas Böhm (started in 2021)

• Marius Keute (2023)
• Fynn Bachmann (2021) → PhD student at the university of Zürich
• Rita González Márquez (2021) → PhD student
• Niklas Böhm (2020) → PhD student

Research

This is a list of [co-]first-/last-author papers grouped by topic; see my Google Scholar page for the complete list ordered chronologically or by citation count. Most papers are open access; for the ones that are not I provide PDFs. Twitter icons link to the respective Twitter threads (tweeprints).

Self-supervised learning

• Nwabufo et al., Self-supervised visualisation of medical image datasets  (arXiv, 2024)
• Böhm et al., Unsupervised visualization of image datasets using contrastive learning  (ICLR 2023) 

Neighbour embeddings

• Lause et al., The art of seeing the elephant in the room: 2D embeddings of single-cell data do make sense  (bioRxiv, 2024) 
• Damrich et al., Visualizing single-cell data with the neighbor embedding spectrum  (bioRxiv, 2024)
• González-Márquez et al., Learning representations of learning representations  (DMLR workshop @ ICLR, 2024) 
• González-Márquez et al., The landscape of biomedical research  ( 🥳 Patterns, 2024) 
• Damrich et al., From t-SNE to UMAP with contrastive learning  (ICLR 2023) 
• Bachmann et al., Wasserstein t-SNE  (ECML/PKDD, 2022) 
• González-Márquez et al., Two-dimensional visualization of large document libraries using t-SNE  (GTRL workshop @ ICLR, 2022) 
• Böhm et al., Attraction-repulsion spectrum in neighbor embeddings  (JMLR, 2022) 
• Kobak & Linderman, Initialization is critical for preserving global data structure in both t-SNE and UMAP  (Nature Biotechnology, 2021)  
• Kobak et al., Heavy-tailed kernels reveal a finer cluster structure in t-SNE visualisations  (ECML/PKDD, 2020) 
• Kobak & Berens, The art of using t-SNE for single-cell transcriptomics  (Nature Communications, 2019) 

Topological data analysis

• Damrich et al., Persistent homology for high-dimensional data based on spectral methods
  (arXiv, 2023)

Statistical methods for transcriptomic and multi-omic data analysis

• Lause et al., Compound models and Pearson residuals for normalization of single-cell RNA-seq data without UMIs  (bioRxiv, 2023)
• Lause et al., Analytic Pearson residuals for normalization of single-cell RNA-seq UMI data  (Genome Biology, 2021) 
• Bernaerts et al., Sparse bottleneck networks for exploratory analysis and visualization of neural Patch-seq data  (arXiv, 2020)
• Kobak et al., Sparse reduced-rank regression for exploratory visualisation of paired multivariate data  (Journal of Royal Statistical Society C, 2021) 

Machine learning & statistical theory

• Greydanus & Kobak, Scaling down deep learning with MNIST-1D  ( 🥳 ICML, 2024)
• Kobak et al., The optimal ridge penalty for real-world high-dimensional data can be zero or negative due to the implicit ridge regularization  (JMLR, 2020) 

Patch-seq data analysis

• Scala* & Kobak* et al., Phenotypic variation in transcriptomic cell types in mouse motor cortex  (Nature, 2021) 
• Scala* & Kobak* et al., Layer 4 of mouse neocortex differs in cell types and circuit organization between sensory areas  (Nature Communications, 2019)

Election forensics

• Kobak et al., Suspect peaks in Russia's “referendum” results  (Significance, 2020) 
• Kobak et al., Putin's peaks: Russian election data revisited  (Significance, 2018)
• Kobak et al., Statistical fingerprints of electoral fraud?  (Significance, 2016)
• Kobak et al., Integer percentages as electoral falsification fingerprints  (The Annals of Applied Statistics, 2016)

Excess mortality / Covid-19 forensics

• Kobak et al., War fatalities in Russia in 2022–2023 estimated via excess male mortality  (SocArXiv, 2023)
• Kobak, Underdispersion: A statistical anomaly in reported Covid data  (Significance, 2022)  
• Karlinsky & Kobak, Tracking excess mortality across countries during the COVID-19 pandemic with the World Mortality Dataset  (eLife, 2021) 
• Kobak, Excess mortality reveals Covid's true toll in Russia  (Significance, 2021) 

Data science

• Kobak et al., Delving into ChatGPT usage in academic writing through excess vocabulary  (arXiv, 2024) 

Previous work and education

In 2013–2016 I was a postdoc in the Machens lab at Champalimaud Institute in Lisbon, working on statistical analysis of electrophyisological population recordings from the cortex.

• Kobak* & Pardo-Vazquez* et al., State-dependent geometry of population activity in rat auditory cortex  (eLife, 2019)
• Kobak* & Brendel* et al., Demixed principal component analysis of neural population data  (eLife, 2016)

In 2007–2012 I did my PhD in the Mehring lab, initially at Freiburg University and later at Imperial College London, working on computational motor control.

• Bashford* & Kobak* et al., Motor skill learning decreases movement variability and increases planning horizon  (Journal of Neurophysiology, 2022)
• Kobak & Mehring, Adaptation paths to novel motor tasks are shaped by prior structure learning  (The Journal of Neuroscience, 2012)

In 2000–2007 I studied computer science (BSc) in St. Petersburg ITMO University and then theoretical physics (MSc) in St. Petersburg State University.

Before that, I attended St. Petersburg Classical Gymnasium #610. I was part-time teaching there computer science and physics in 2002–2006 while studying in university. In 2004, together with a friend, I made a website 610.ru that is still online (with minor changes).

Reviewing

Year  Reviews
-------------
2019  1
2020  15
2021  20
2022  22
2023  28
2024  14

Venue              Reviews (>1)
-------------------------------
NeurIPS            24
ECML               12
ICML               12
ICLR               11
TMLR               6
Bioinformatics     4
Genome Biology     4
AISTATS            3
JMLR               3
Nature Comm.       2
Nature Methods     2
PLoS Comp Bio      2
Political Analysis 2

Review lengths in kB
--------------------
 0 | 
 1 | ................
 2 | ...............
 3 | .......................
 4 | .................
 5 | ...........
 6 | ............
 7 | ...
 8 | ..
 9 | .

Last updated: Jul 11, 2024