Sitemap
A list of all the posts and pages found on the site. For you robots out there is an XML version available for digesting as well.
Pages
Books
by
CV
by
Markdown
by
Page Archive
by
Misc
by
Publications
by
Sitemap
by
Talk map
by
Teaching
by
Blog posts
by
Posts
Published:
This post will show up by default. To disable scheduling of future posts, edit
config.yml
and set future: false
.Published:
This is a sample blog post. Lorem ipsum I can’t remember the rest of lorem ipsum and don’t have an internet connection right now. Testing testing testing this blog post. Blog posts are cool.
Published:
This is a sample blog post. Lorem ipsum I can’t remember the rest of lorem ipsum and don’t have an internet connection right now. Testing testing testing this blog post. Blog posts are cool.
Published:
This is a sample blog post. Lorem ipsum I can’t remember the rest of lorem ipsum and don’t have an internet connection right now. Testing testing testing this blog post. Blog posts are cool.
Published:
This is a sample blog post. Lorem ipsum I can’t remember the rest of lorem ipsum and don’t have an internet connection right now. Testing testing testing this blog post. Blog posts are cool.
books
Analiza danych z programem R
by
Published:
An academic textbook describing estimation and testing topics for linear models with fixed effects, random effects and mixed effects. The theoretical introduction is complemented by numerous examples for one-way and multivariate ANOVA, one and multiple random components. The examples focus on biological and medical applications and are based on real analyses of real data.
Eseje o sztuce wizualizacji danych
by
Published:
Discover! Reveal! Explain! These three roles can be fulfilled by good statistical graphics. Good means understandable, faithful to the data, aesthetic. How to create such graphics? A collection of essays on the art of displaying data systematises knowledge useful in designing and producing good data visualisations. It is not easy. On the one hand, we can fall into the trap of a colourful mush full of numbers, which is sometimes proudly called infographics. On the other hand, we can fall into the trap of graphics that perfectly reproduce the complexity of numbers, and thus completely incomprehensible. Somewhere in the middle is a graphic that explains, that informs, that is aesthetically pleasing and informative.
Published:
The Guide to the R package was the first published Polish book focused on the R language. The current fourth edition consists of four parts: Basics of using R (+tidyverse, shiny, knitr and other goodies), Programming in R (object-oriented, package development, class system), Statistics with R (statistical tests, models, exploration techniques) and Visualization with R (graphics, lattice and ggplot2 packages).
Chaos Game with examples in R, Python and Julia.
by
Published:
Are you curious about fractals? The Chaos Game is the book for you. You will learn the mathematical basis behind these figures, find out what algorithm can be used to code them, write code in your favourite programming language (Python, R, Julia?) and also explore the bibliographies of three mathematicians associated with the development of mathematics around these shapes. This is the next book in the Beta Bit series for anyone interested in computational mathematics and data analysis.
The Hitchhiker’s Guide to Responsible Machine Learning
by
Published:
A one-of-a-kind 52-page story about responsible machine learning. Beta and Bit use decision trees, random forests, and AutoML tools to build a risk model after a covid infection, and then use explainable artificial intelligence tools to analyze the behavior of that model. The description of the data analysis process is intertwined with descriptions of ML tools and code snippets. All examples are fully reproducible!
Published:
How can you create good plots? Good ones, meaning ones that are viewed with pleasure, that are informative, that are understood by a wide audience but that still can be appreciated by connoisseurs.
Published:
A step-by-step introduction to the most important methods of explainable machine learning (XAI). Learn the intuitions, mathematical foundations and application examples of LIME, SHAP, Break Down, Partial Dependece, Permutational Variable Importance, Accumulated Local Effects and other popular techniques. Methods are enhanced with examples in R and Python using the DALEX library to explore any predictive model.
portfolio
2024-12-15 [ENG]: How I use GenAI at work: MidJourney
by
Generative AI is transforming how we work, but the extent of this change remains uncertain. Adopting these tools in academia presents unique challenges, yet this shift is inevitable. This blog series documents how I use selected GenAI tools in late 2024, offering insights that might inspire others in academia or serve as a personal reference for the early days of this transformation.
In this post, I delve into MidJourney, a versatile tool for generating creative visuals. I explain why I prefer it over similar tools, describe how I use it (e.g., for presentations, comics, and videos), and outline its strengths and ethical considerations. While MidJourney excels in visual storytelling, I caution against using it for research graphics or imitating contemporary artists’ styles. How do you use MidJourney? Share your thoughts in the comments!

2024-12-15 [PL]: Jak wykorzystuję GenAI w pracy: MidJourney
by
Generatywna AI zmienia sposób, w jaki pracujemy, ale w jakim zakresie - to się dopiero okaże. Wdrażanie narzędzi GenAI w środowisku akademickim wiąże się z wyjątkowymi wyzwaniami, jednak ta zmiana jest nieunikniona. Planowana krótka seria blogów dokumentuje, jak korzystam z wybranych narzędzi GenAI pod koniec 2024 roku, oferując spostrzeżenia, które mogą się przydać innym osobom w środowisku akademickim.
W tym wpisie omawiam MidJourney, wszechstronne narzędzie do generowania kreatywnych grafik. Wyjaśniam, dlaczego je preferuję w porównaniu do podobnych narzędzi, opisuję sposoby, w jakie je wykorzystuję (np. do prezentacji, komiksów i filmów). Chociaż MidJourney doskonale sprawdza się w wizualnym opowiadaniu historii, przestrzegam przed używaniem go do grafik naukowych lub imitowania stylów współczesnych artystów. A jak Ty korzystasz z MidJourney? Podziel się swoją opinią w komentarzach!

2024-12-20 [ENG]: How I use GenAI at work: NotebookLM
by
In this post, I discuss NotebookLM, a generative AI tool designed for interactive exploration of documents. We go over its key features, including the ability to ask questions in natural language, generate audio summaries, and organize responses into notes, making it a versatile tool for academic and professional use. The post emphasizes three primary applications: refining one’s own papers, creating podcasts summarizing articles, and brainstorming teaching strategies like storytelling and generating examples. It also cautions against using NotebookLM for reviewing research papers or handling sensitive documents due to privacy and reliability concerns. And how do you use NotebookLM? Share your thoughts in the comments!

2024-12-20 [PL]: Jak wykorzystuję GenAI w pracy: NotebookLM
by
W tym wpisie omawiam NotebookLM, narzędzie generatywnej AI zaprojektowane do interaktywnej eksploracji dokumentów. Opisuję jego funkcje, które najczęściej wykorzystuję, takie jak zadawanie pytań w języku naturalnym, generowanie audio podsumowań i organizowanie odpowiedzi w formie notatek. Te funkcje czynią NotebookLM wszechstronnym narzędziem do zastosowań na uczelni. Ten wpis koncentruje się na trzech głównych zastosowaniach: dopracowywaniu własnych artykułów, tworzeniu podcastów podsumowujących prace naukowe oraz eksploracja kreatywnych sposóbów przedstawiania złożonych tematów. Ten wpis wskazuje też obszary, w których odradzam używanie NotebookLM, takie jak recenzowanie prac naukowych czy pracyaz wrażliwymi dokumentami. A jak Ty korzystasz z NotebookLM? Podziel się swoimi przemyśleniami w komentarzach!

2024-12-21 [ENG]: How I use GenAI at work: HeyGen
by
HeyGen is a powerful generative AI tool that transforms how we create video content by simplifying the process of generating professional-quality videos with customizable avatars. In this blog, I share my experience with HeyGen and its applications in academic and creative workflows. From creating concise video summaries of complex projects to generating multilingual content effortlessly, HeyGen has proven to be a game-changer. Its excellent text-to-speech capabilities, customizable avatars, and efficient workflow streamline video production, saving time and enhancing accessibility. While there are some limitations, like the inability to use animated backgrounds, HeyGen’s potential in education, outreach, and beyond is immense. And how do you use HeyGen? Share your thoughts in the comments!

2024-12-21 [PL]: Jak wykorzystuję GenAI w pracy: HeyGen
by
HeyGen to potężne narzędzie generatywnej sztucznej inteligencji, które zmienia sposób, w jaki tworzymy treści wideo, upraszczając proces generowania profesjonalnych filmów z personalizowanymi awatarami. W tym wpisie dzielę się swoimi doświadczeniami z HeyGen i jego zastosowaniami w pracy akademickiej oraz kreatywnej. Od tworzenia zwięzłych podsumowań wideo skomplikowanych projektów po łatwe generowanie treści wielojęzycznych. Dla mnie HeyGen okazał się przełomowym rozwiązaniem. Jego doskonałe możliwości konwersji tekstu na mowę, personalizowane awatary i prosty interface usprawniają produkcję wideo, oszczędzając czas i zwiększając dostępność geenrowania nowych treści. Choć ma pewne ograniczenia (o czym poniżej) to potencjał HeyGen w edukacji, popularyzacji wiedzy i innych dziedzinach jest ogromny. A jak Ty wykorzystujesz HeyGen? Podziel się swoją opinią w komentarzach!

2024-12-22 [ENG]: How I use GenAI at work: ChatGPT
by
This blog focusing on the versatile applications of ChatGPT in research, teaching, and PR. It highlights ChatGPT’s ability to streamline brainstorming, simulate discussions on academic topics, assist in translations, and support scriptwriting in various programming languages. While acknowledging its strengths, such as improving text consistency and facilitating creative problem-solving. Also we identify limitations. These include ChatGPT’s inability to critically evaluate tasks, generate reliable academic references, or review non-public documents due to privacy concerns. And how do you use ChatGPT? Share your thoughts in the comments!

2024-12-22 [PL]: Jak korzystam z GenAI w pracy: ChatGPT
by
Ten blog skupia się na wszechstronnych zastosowaniach ChatGPT w badaniach naukowych, nauczaniu i PR. Podkreśla zdolność ChatGPT do usprawniania burzy mózgów, symulowania dyskusji na tematy akademickie, wspierania tłumaczeń oraz tworzenia skryptów w różnych językach programowania. Wymieniam zalety z których korzystam, takie jak poprawa spójności tekstów i wspieranie kreatywnego rozwiązywania problemów, ale też wskazuję na ograniczenia, m.in. brak krytycznej oceny zadań, generowania wiarygodnych źródeł naukowych czy recenzji niepublicznych dokumentów ze względu na kwestie prywatności. A jak Ty korzystasz z ChatGPT? Podziel się w komentarzach!

publications
An international comparison of students’ ability to endure fatigue and maintain motivation during a low-stakes test
by Francesca Borgonovi, Przemyslaw Biecek
published in Learning and Individual Differences, 2016 #PISA #policy
This paper investigates academic endurance—students’ ability to sustain performance throughout low-stakes cognitive tests—using PISA data from over 450,000 students across 60+ countries. The study finds significant endurance disparities across countries and subgroups, with sharper declines in performance among boys, socio-economically disadvantaged students, and during reading assessments, particularly influenced by question format.
Show Me Your Model: tools for visualisation of statistical models
by Przemyslaw Biecek
published in UseR, 2017 #R #software
This talk explores the evolving landscape of statistical model visualization, highlighting the diversity of existing R packages and the lack of a unified structure across them. It advocates for a consistent grammar of model visualization—akin to ggplot2 grammar of graphics—enabled by tools like broom and supported by emerging theoretical frameworks.
The Merging Path Plot: adaptive fusing of k-groups with likelihood-based model selection
by Agnieszka Sitko, Przemyslaw Biecek
published in arxiv, 2017 #R #software
This article introduces the Merging Path Plot and the factorMerger R package, which provide an intuitive method to explore and visualize dissimilarities among k-groups after rejecting the global null hypothesis. Unlike traditional pairwise post hoc tests, this approach leverages Likelihood Ratio Tests to effectively summarize group differences, especially when dealing with many groups.
intsvy: An R Package for Analyzing International Large-Scale Assessment Data
by Daniel H. Caro, Przemyslaw Biecek
published in Journal of Statistical Software, 2017 #R #software
This paper presents intsvy, an R package designed to streamline the import, analysis, and visualization of international large-scale assessment data such as PISA, TIMSS, and PIRLS. The package accounts for complex survey designs and test structures, offering robust statistical and graphical tools tailored for cross-national education research.
archivist: An R Package for Managing, Recording and Restoring Data Analysis Results
by Przemyslaw Biecek, Marcin Kosiński
published in Journal of Statistical Software, 2017 #R #software
This article introduces archivist, an R package that enhances reproducible research by archiving R objects along with their metadata, creation context, and relationships. It enables efficient storage, retrieval, verification, and sharing of analytical results, opening new possibilities for traceable and interactive data analysis workflows.
DALEX: Explainers for Complex Predictive Models in R
by Przemyslaw Biecek
published in Journal of Machine Learning Research, 2018 #R #software
This paper presents a unified framework of model-agnostic explainers for interpreting complex predictive models, such as neural networks and ensembles, regardless of their internal structure. Implemented in the R package DALEX, these explainers support understanding and comparing model behavior through standardized techniques for decomposing predictions, assessing variable importance, and evaluating model performance.
Explanations of Model Predictions with live and breakDown Packages
by Mateusz Staniak, Przemyslaw Biecek
published in The R Journal, 2018 #R #software
This paper introduces two R packages, live and breakDown, for attributing predictions from complex black-box models to input features, enhancing interpretability in predictive modeling. The authors compare these new approaches with existing state-of-the-art tools like lime and iml, demonstrating alternative methods for local explanation and model understanding.
survxai: an R package for structure-agnostic explanations of survival models
by Aleksandra Grudziaz, Alicja Gosiewska, Przemyslaw Biecek
published in The Journal of Open Source Software, 2018 #R #software
This paper introduces survxai, an R package that provides local and global, structure-agnostic explanations for complex survival models, addressing a gap left by existing model-agnostic tools focused mainly on regression and classification. By enabling interpretability of survival functions across models like Cox and survival random forests, survxai supports trust, understanding, and improved performance in high-stakes applications such as medicine and churn analysis.
auditor: an R Package for Model-Agnostic Visual Validation and Diagnostics
by Alicja Gosiewska, Przemyslaw Biecek
published in The R Journal, 2019 #R #software
This paper presents auditor, an R package offering model-agnostic tools for auditing predictive models by assessing performance, goodness of fit, residual behavior, and identifying outliers or influential points. Through diagnostic scores and visualizations, auditor enables robust, flexible validation across diverse model types, supporting real-world reliability beyond test accuracy.
The Landscape of R Packages for Automated Exploratory Data Analysis
by Mateusz Staniak, Przemyslaw Biecek
published in The R Journal, 2019 #R #software
This paper presents a systematic review of fifteen popular R packages for Automated Exploratory Data Analysis (autoEDA), highlighting how current tools automate key aspects of EDA such as data cleaning, validation, and feature engineering. The review identifies both the strengths and limitations of existing solutions, offering guidance for future developments in accelerating and enhancing the data exploration process.
modelStudio: Interactive Studio with Explanations for ML Predictive Models
by Hubert Baniecki, Przemyslaw Biecek
published in The Journal of Open Source Software, 2019 #R #software
This paper discusses the growing need for automating the explanation of machine learning models, paralleling trends in AutoML and AutoEDA, due to the complexity and time demands of interpretability tasks. While tools like DALEX, lime, and shap support interpretability, the paper critiques current automation solutions like modelDown for focusing only on global explanations and lacking interactivity, highlighting the need for more dynamic and accessible XAI tools.
Named Entity Recognition - Is There a Glass Ceiling?
by Tomasz Stanislawek, Anna Wróblewska, Alicja Wójcicka, Daniel Ziembicki, Przemyslaw Biecek
published in Conference on Computational Natural Language Learning (CoNLL), 2019 #NLP #benchmark
This paper provides an in-depth error analysis of leading Named Entity Recognition (NER) models, including Stanford, CMU, FLAIR, ELMO, and BERT, revealing their strengths, weaknesses, and shared limitations. It also proposes novel techniques to enhance annotation, training, and evaluation processes, aiming to overcome persistent challenges and push beyond the current performance ceiling in NER.
Models in the Wild: On Corruption Robustness of Neural NLP Systems
by Barbara Rychalska, Dominika Basaj, Alicja Gosiewska, Przemyslaw Biecek
published in International Conference on Neural Information Processing (ICONIP), 2019 #NLP #benchmark
This paper introduces WildNLP, a framework for evaluating the robustness of NLP models against natural text corruptions like typos and keyboard errors across four major tasks. The study reveals that high-performing models often lack robustness, though modern embedding techniques and adversarial training can improve stability; the framework is publicly available for further research.
Explainable Machine Learning for Modeling of Early Postoperative Mortality in Lung Cancer
by Katarzyna Kobylińska, Tomasz Mikołajczyk, Mariusz Adamek, Tadeusz Orłowski, Przemyslaw Biecek
published in Artificial Intelligence in Medicine, 2020 #Surv-XAI #BioMed
This study applies explainable machine learning techniques to predictive models for early postoperative mortality in lung cancer patients, using data from the Domestic Lung Cancer Database. It demonstrates how model interpretability can enhance clinical decision-making by combining data-driven insights with domain expertise, addressing critical risks associated with black-box models in high-stakes medical contexts.
KRAB ZNF explorer—the online tool for the exploration of the transcriptomic profiles of KRAB-ZNF factors in The Cancer Genome Atlas
by Rafał Cylwa, Kornel Kiełczewski, Marta Machnik, Urszula Oleksiewicz , Przemyslaw Biecek
published in Bioinformatics, 2020 #BioMed #software
The KRAB ZNF Explorer is a web-based tool designed to analyze mRNA expression patterns of KRAB-ZNF transcription factors using data from The Cancer Genome Atlas. It enables comparative analyses between normal and cancer tissues, explores associations with pathological features and DNA methylation, and includes survival analysis and isoform-level expression profiling.
What Would You Ask the Machine Learning Model? Identification of User Needs for Model Explanations Based on Human-Model Conversations
by Michał Kuźba, Przemyslaw Biecek
published in European Conference on Machine Learning (ECML), 2021 #model-exploration #Interactive-XAI
This paper introduces dr_ant, a conversational system designed to explore what types of questions human users naturally ask when interacting with machine learning models, using a Titanic survival prediction model as a case study. By analyzing over 1,000 dialogues, the study reveals common user information needs, offering a novel approach to aligning explainable AI methods with real human expectations and inquiry patterns.
Explanatory Model Analysis. Explore, Explain, and Examine Predictive Models
by Przemyslaw Biecek, Tomasz Burzykowski
published in Chapman and Hall/CRC, 2021 #model-exploration #Python #software
This book introduces Explanatory Model Analysis (EMA), a comprehensive suite of model-agnostic methods for exploring, explaining, and examining predictive models to improve their reliability and transparency. Addressing the core bottleneck in modern machine learning—model interpretability—it offers practical tools and real-world applications for both classification and regression tasks.
Transparency, auditability, and explainability of machine learning models in credit scoring
by Michael Buecker, Gero Szepannek, Alicja Gosiewska, Przemyslaw Biecek
published in Journal of the Operational Research Society, 2021 #policy #credit
This article proposes a comprehensive framework to make black-box machine learning models in credit scoring transparent, auditable, and explainable, addressing the industrys regulatory demands for interpretability. Through a real-world case study, it demonstrates that advanced ML models can achieve interpretability comparable to traditional scorecards while maintaining superior predictive performance.
dalex: Responsible Machine Learning with Interactive Explainability and Fairness in Python
by Hubert Baniecki, Wojciech Kretowicz, Piotr Piątyszek, Jakub Wiśniewski, Przemyslaw Biecek
published in Journal of Machine Learning Research, 2021 #Python #software
This paper introduces dalex, a Python package designed to address the challenges of model opaqueness by providing a unified, model-agnostic interface for explainability and fairness in machine learning. Aimed at promoting responsible AI development, dalex supports validation, transparency, and continuous monitoring, aligning with both scientific demands and emerging regulatory requirements.
Kleister: Key Information Extraction Datasets Involving Long Documents with Complex Layouts
by Tomasz Stanisławek, Filip Graliński, Anna Wróblewska, Dawid Lipiński, Agnieszka Kaliska, Paulina Rosalska, Bartosz Topolski, Przemyslaw Biecek
published in International Conference on Document Analysis and Recognition (ICDAR), 2021 #NLP #benchmark
This paper introduces two challenging new datasets—Kleister NDA and Kleister Charity—for advancing Key Information Extraction (KIE) in NLP, involving complex formal documents with both scanned and digital formats. By benchmarking state-of-the-art models and sharing these datasets, the authors aim to drive progress in extracting structured information using both textual and layout features.
Checklist for responsible deep learning modeling of medical images based on COVID-19 detection studies
by Weronika Hryniewska, Przemysław Bombiński, Patryk Szatkowski, Paulina Tomaszewska, Artur Przelaskowski, Przemysław Biecek
published in Pattern Recognition, 2021 #BioMed #benchmark
This paper conducts a systematic review of deep learning models developed for COVID-19 detection, highlighting common errors in data handling, model development, and explainability due to insufficient radiographic domain understanding. Combining insights from radiologists and ML engineers, the authors propose a checklist of essential criteria for building reliable diagnostic models.
Simpler is better: Lifting interpretability-performance trade-off via automated feature engineering
by Alicja Gosiewska, Anna Kozak, Przemysław Biecek
published in Decision Support Systems, 2021 #AutoML #benchmark
This paper presents a framework that transforms complex black-box models into interpretable glass-box models by extracting informative features from the former, enabling simplification without sacrificing accuracy. Through large-scale benchmarking and real-world application, the study demonstrates that simpler models can match or exceed the performance of complex ones, challenging the belief that predictive accuracy necessitates opacity.
Towards Explainable Meta-learning
by Katarzyna Woźnica, Przemyslaw Biecek
published in European Conference on Machine Learning (ECML), 2022 #AutoML #benchmark
This paper pioneers the application of eXplainable AI (XAI) techniques to meta-learning by using post-hoc explainability to analyze black-box surrogate models and uncover the influence of meta-features on model performance. It highlights the need for interpretability in building next-generation meta-models and represents a novel step toward more transparent and insightful meta-learning systems.
Explainable AI Methods - A Brief Overview
by Andreas Holzinger, Anna Saranti, Christoph Molnar, Przemyslaw Biecek, Wojciech Samek
published in xxAI - Beyond Explainable AI, 2022 #Survey #post-hoc
This article provides a concise overview of 17 prominent methods in Explainable AI (xAI), aiming to familiarize beginners—particularly application engineers and data scientists—with the current state of the field. It summarizes techniques ranging from model-agnostic approaches like LIME and SHAP to specialized methods for visual, textual, and causal explanations.
fairmodels: a Flexible Tool for Bias Detection, Visualization, and Mitigation in Binary Classification Models
by Jakub Wiśniewski, Przemyslaw Biecek
published in The R Journal, 2022 #R #software
This article introduces fairmodels, an R package designed to assess and mitigate bias in binary classification models through a model-agnostic framework. It provides comprehensive tools for fairness validation, visualization, and comparison, enabling responsible development of predictive systems by addressing potential discrimination in historical data.
LIMEcraft: handcrafted superpixel selection and inspection for Visual eXplanations
by Weronika Hryniewska, Adrianna Grudzień, Przemyslaw Biecek
published in Machine Learning, 2022 #BioMed #Interactive-XAI
This paper presents LIMEcraft, an interactive tool designed to improve the interpretability and safety of deep learning models by enabling users to select and examine semantically meaningful image regions during explanation. By going beyond static saliency maps, LIMEcraft enhances bias detection and model fairness assessment, offering a more thorough understanding of image-based predictions.
SurvSHAP(t): Time-dependent explanations of machine learning survival models
by Mateusz Krzyziński, Mikołaj Spytek, Hubert Baniecki, Przemysław Biecek
published in Knowledge-Based Systems, 2023 #BioMed #shap
This paper introduces SurvSHAP(t), the first time-dependent, model-agnostic explanation method tailored for interpreting survival function predictions from complex black-box models. Built on the widely adopted SHAP framework, SurvSHAP(t) improves interpretability in precision diagnostics by identifying time-varying effects of variables and outperforming existing methods like SurvLIME in explanatory power.
Fooling Partial Dependence via Data Poisoning
by Hubert Baniecki, Wojciech Kretowicz, Przemyslaw Biecek
published in European Conference on Machine Learning (ECML), 2023 #Adversarial-XAI #model-exploration
This paper reveals that Partial Dependence explanations for predictive models can be adversarially manipulated through data poisoning using genetic and gradient-based algorithms, raising serious concerns about their reliability in high-stakes domains. It introduces the first known use of genetic algorithms to fool model explanations in a transferable, model- and explanation-agnostic way, highlighting critical vulnerabilities in post-hoc interpretability methods.
Towards Evaluating Explanations of Vision Transformers for Medical Imaging
by Piotr Komorowski, Hubert Baniecki, Przemyslaw Biecek
published in Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2023 #Red-Teaming #BioMed
This paper evaluates the interpretability of Vision Transformers (ViTs) in medical imaging by comparing explanation methods on chest X-ray classification tasks. It introduces criteria like faithfulness, sensitivity, and complexity, and finds that Layerwise Relevance Propagation outperforms other techniques, offering more accurate and reliable insights into ViT decision-making.
Consolidated learning: a domain-specific model-free optimization strategy with validation on metaMIMIC benchmarks
by Katarzyna Woźnica, Mateusz Grzyb, Zuzanna Trafas, Przemyslaw Biecek
published in Machine Learning, 2023 #AutoML #BioMed
This paper introduces consolidated learning, a novel approach to hyperparameter tuning that optimizes across multiple similar tasks rather than focusing on single-task performance, addressing practical challenges in real-world model development. By leveraging static portfolios of hyperparameter configurations and effective transfer across tasks, the method demonstrates strong anytime performance and efficiency, validated on XGBoost using the metaMIMIC benchmark derived from MIMIC-IV medical data.
Generative AI models should include detection mechanisms as a condition for public release
by Alistair Knott, Dino Pedreschi, Raja Chatila, Tapabrata Chakraborti, Susan Leavy, Ricardo Baeza-Yates, David Eyers, Andrew Trotman, Paul D. Teal, Przemyslaw Biecek, Stuart Russell, Yoshua Bengio
published in Ethics and Information Technology , 2023 #policy #GenAI
This paper proposes a regulatory principle requiring developers of publicly released foundation models to provide a reliable, publicly accessible detection tool for identifying AI-generated content. The authors argue that such a mechanism is technically feasible and vital for mitigating risks associated with generative AI, while outlining potential design options and areas needing further policy and research input.
survex: an R package for explaining machine learning survival models
by Mikołaj Spytek, Mateusz Krzyzinski, Sophie Hanna Langbein, Hubert Baniecki, Marvin N. Wright, Przemysław Biecek
published in Bioinformatics, 2023 #BioMed #shap
This paper presents survex, an R package offering a unified framework for explaining survival models using explainable AI techniques, addressing the interpretability gap in machine learning-based survival analysis. By revealing variable effects, importances, and model decision rationale, survex enhances model transparency, reliability, and trust—especially crucial in sensitive domains like healthcare and biomedical research.
Protect our environment from information overload
by Janusz A. Hołyst, Philipp Mayr, Michael Thelwall, Ingo Frommholz, Shlomo Havlin, Alon Sela, Yoed N. Kenett, Denis Helic, Aljoša Rehar, Sebastijan R. Maček, Przemysław Kazienko, Tomasz Kajdanowicz, Przemyslaw Biecek, Boleslaw K. Szymanski, Julian Sienkiewicz
published in Nature Human Behaviour, 2024 #policy #information-overload
This article argues that information overload (IOL) should be treated as a critical environmental factor, warranting deeper research into its cognitive, emotional, and contextual impacts. It highlights IOL as a complex, multidimensional condition that impairs task performance and well-being, urging the development of strategies for managing our increasingly saturated information environment.
The grammar of interactive explanatory model analysis
by Hubert Baniecki, Dariusz Parzych, Przemyslaw Biecek
published in Data Mining and Knowledge Discovery , 2024 #model-exploration #Interactive-XAI
This paper introduces Interactive Explanatory Model Analysis (IEMA), a framework that emphasizes the need for sequential, multi-method interpretation to address the limitations of isolated explanations in machine learning. By formalizing the grammar of IEMA and implementing it in a user-centered open-source tool, the authors demonstrate through user studies that combining diverse explanatory methods enhances human understanding, confidence, and decision-making accuracy.
Red Teaming Models for Hyperspectral Image Analysis Using Explainable AI
by Vladimir Zaigrajew, Hubert Baniecki, Lukasz Tulczyjew, Agata M. Wijata, Jakub Nalepa, Nicolas Longépé, Przemyslaw Biecek
published in ICLR Machine Learning for Remote Sensing (ML4RS) Workshop, 2024 #Red-Teaming #Hyperspectral
This paper presents a red-teaming methodology for evaluating machine learning models in remote sensing, specifically targeting hyperspectral image analysis in the HYPERVIEW challenge. By applying explainable AI techniques, the authors expose critical model weaknesses, propose a more efficient alternative using only 1% of input features with minimal performance loss, and introduce a domain-informed visualization approach tailored for hyperspectral data interpretation.
Performance Is Not Enough: The Story Told by a Rashomon Quartet
by Przemyslaw Biecek, Hubert Baniecki, Mateusz Krzyziński, Dianne Cook
published in Journal of Computational and Graphical Statistics, 2024 #Rashomon #model-exploration
This article introduces the Rashomon Quartet—four models with nearly identical predictive performance yet markedly different explanations—highlighting the limitations of relying solely on accuracy in supervised learning. Inspired by Anscombe’s quartet, it underscores the importance of visual model comparison to uncover divergent data interpretations among equally effective models.
A novel radiomics approach for predicting TACE outcomes in hepatocellular carcinoma patients using deep learning for multi-organ segmentation
by Krzysztof Bartnik, Mateusz Krzyziński, Tomasz Bartczak, Krzysztof Korzeniowski, Krzysztof Lamparski, Tadeusz Wróblewski, Michał Grąt, Wacław Hołówko, Katarzyna Mech, Joanna Lisowska, Magdalena Januszewicz, Przemysław Biecek
published in Scientific Reports, 2024 #Segmentation-XAI #BioMed
This study introduces an automated radiomics-based machine learning approach for predicting transarterial chemoembolization (TACE) outcomes in hepatocellular carcinoma patients, using features from multiple organ volumes of interest (VOIs) extracted from pre-TACE CT scans. The method outperforms clinical models in predicting progression-free survival and highlights the prognostic value of non-tumoral VOIs, offering a scalable and radiologist-free alternative for treatment outcome prediction.
Red-Teaming Segment Anything Model
by Krzysztof Jankowski, Bartlomiej Sobieski, Mateusz Kwiatkowski, Jakub Szulc, Michał Janik, Hubert Baniecki
published in Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2024 #Red-Teaming #SAM
This paper conducts a comprehensive red-teaming analysis of the Segment Anything Model (SAM), exposing vulnerabilities in segmentation performance under style transformations, privacy-related misuse, and adversarial attacks. The findings, including the introduction of the Focused Iterative Gradient Attack (FIGA), highlight significant safety concerns and underscore the need for improved robustness and ethical safeguards in foundation models for image segmentation.
Position: Explain to Question not to Justify
by Przemyslaw Biecek, Wojciech Samek
published in International Conference on Machine Learning (ICML), 2024 #RED-XAI #BLUE-XAI
This position paper introduces a conceptual division in XAI between BLUE XAI (human/value-oriented) and RED XAI (model/validation-oriented), emphasizing the latter’s critical yet underexplored role in ensuring AI safety. The authors advocate for intensified research in RED XAI to enable deeper model interrogation, bug detection, and knowledge extraction, outlining key challenges and opportunities in this emerging direction.
SRFAMap: A Method for Mapping Integrated Gradients of a CNN Trained with Statistical Radiomic Features to Medical Image Saliency Maps
by Oleksandr Davydko, Vladimir Pavlov, Przemyslaw Biecek, Luca Longo
published in World Conference on Explainable Artificial Intelligence, 2024 #post-hoc #BioMed
This paper introduces SRFAMap, a novel method that translates integrated gradients of statistical radiomic features into saliency maps, enabling visual interpretation of feature contributions in medical image classification where gradients are otherwise inaccessible. Applied to chest X-rays analyzed by a ResNet-50 model, SRFAMap facilitates diagnostics by highlighting lesion-relevant regions with faithful and statistically significant visual attributions.
Adversarial attacks and defenses in explainable artificial intelligence: A survey
by Hubert Baniecki, Przemyslaw Biecek
published in Information Fusion, 2024 #Adversarial-XAI #Secure-XAI
This survey highlights the vulnerabilities of explainable AI (XAI) methods to adversarial manipulation, raising concerns about their reliability in high-stakes applications. It introduces a unified taxonomy bridging adversarial machine learning and XAI, reviews attack and defense strategies, and outlines future directions for developing more robust and secure explanation techniques.
AI content detection in the emerging information ecosystem: new obligations for media and tech companies
by Alistair Knott, Dino Pedreschi, Toshiya Jitsuzumi, Susan Leavy, David Eyers, Tapabrata Chakraborti, Andrew Trotman, Sundar Sundareswaran, Ricardo Baeza-Yates, Przemyslaw Biecek, Adrian Weller, Paul D. Teal, Subhadip Basu, Mehmet Haklidir, Virginia Morini, Stuart Russell, Yoshua Bengio
published in Ethics and Information Technology, 2024 #GenAI #policy
This paper explores emerging regulatory obligations—driven by the EU AI Act and U.S. Executive Order—that require AI content providers to support reliable detection of AI-generated content, framing it as essential for societal trust and resilience. The authors argue that these developments create a new adversarial landscape, prompting policymakers to impose duties on media, search platforms, and the broader tech ecosystem to ensure the effective deployment and governance of AI-content detection mechanisms.
WAW-TACE: A Hepatocellular Carcinoma Multiphase CT Dataset with Segmentations, Radiomics Features, and Clinical Data
by Krzysztof Bartnik, Tomasz Bartczak, Mateusz Krzyziński, Krzysztof Korzeniowski, Krzysztof Lamparski, Piotr Węgrzyn, Eric Lam, Mateusz Bartkowiak, Tadeusz Wróblewski, Katarzyna Mech, Magdalena Januszewicz, Przemysław Biecek
published in Radiology: Artificial Intelligence, 2024 #Segmentation-XAI #BioMed
The WAW-TACE dataset provides comprehensive multimodal data from 233 treatment-naive hepatocellular carcinoma patients undergoing transarterial chemoembolization, including multiphase CT scans, 377 handcrafted tumor masks, automated organ segmentations, radiomics features, and detailed clinical information. It is publicly available for research use at Zenodo.
Swin SMT: Global Sequential Modeling for Enhancing 3D Medical Image Segmentation
by Maciej Chrabaszcz, Szymon Plotka, Przemyslaw Biecek
published in Medical Image Computing and Computer Assisted Intervention (MICCAI), 2024 #GenXAI #BioMed
This paper presents Swin Soft Mixture Transformer (Swin SMT), a novel architecture that enhances 3D medical image segmentation by integrating a Soft Mixture-of-Experts mechanism into Swin UNETR to better capture diverse long-range dependencies. Evaluated on the TotalSegmentator-V2 dataset, Swin SMT achieves state-of-the-art performance with an average Dice Similarity Coefficient of 85.09%, while maintaining efficiency in both training and inference.
CNN-Based Explanation Ensembling for Dataset, Representation and Explanations Evaluation
by Weronika Hryniewska, Luca Longo, Przemyslaw Biecek
published in World Conference on Explainable Artificial Intelligence, 2024 #Ensemble-XAI #post-hoc
This paper explores the ensembling of explanations from deep classification models to produce more coherent and reliable insights into model behavior, particularly in high-stakes domains. The proposed method not only improves explanation quality-demonstrated through enhanced Localization and Faithfulness metrics-but also aids in detecting class under-representation and reducing sensitive information by transforming input images into their explanation-based representations.
Resistance Against Manipulative AI: Key Factors and Possible Actions (RAMAI)
by Piotr Wilczyński, Wiktoria Mieleszczenko-Kowszewicz, Przemyslaw Biecek
published in European Conference on Artificial Intelligence (ECAI), 2024 #LLM #policy
This paper investigates the potential of large language models (LLMs) to manipulate human decisions by analyzing both user susceptibility and LLM persuasion strategies through controlled experiments. It proposes two mitigation strategies: promoting AI literacy to reduce long-term risks and introducing a “Manipulation Fuse” classifier to detect manipulative content in real-time.
Big Tech influence over AI research revisited: Memetic analysis of attribution of ideas to affiliation
by Stanisław Giziński, Paulina Kaczyńska, Hubert Ruczyński, Emilia Wiśnios, Bartosz Pieliński, Przemysław Biecek, Julian Sienkiewicz
published in Journal of Informetrics, 2024 #policy #Graph-XAI
This paper critically examines the influence of Big Tech in AI research, moving beyond publication volume to analyze the spread of ideas using memetic and network analysis across large-scale datasets. The findings reveal a more nuanced landscape, showing that while Big Tech plays a significant role, the most influential AI ideas often emerge from collaborations between Big Tech and academia, challenging oversimplified narratives of dominance.
Exploration of the Rashomon Set Assists Trustworthy Explanations for Medical Data
by Katarzyna Kobylińska, Mateusz Krzyziński, Rafał Machowicz, Mariusz Adamek, Przemyslaw Biecek
published in IEEE Journal of Biomedical and Health Informatics, 2024 #Rashomon #BioMed
This paper proposes Rashomon_DETECT, a novel algorithm for identifying behaviorally distinct models within the Rashomon set, addressing limitations of conventional single-model selection in high-stakes domains like medicine. By introducing the Profile Disparity Index (PDI) and leveraging XAI techniques, the method enables deeper analysis of model variability, enhancing trustworthiness and interpretability in predictive modeling.
Position: Do Not Explain Vision Models Without Context
by Paulina Tomaszewska, Przemyslaw Biecek
published in International Conference on Machine Learning (ICML), 2025 #policy #BioMed
This paper critiques current explanation methods for computer vision models for neglecting spatial context, which is essential for accurate interpretation of visual scenes. It highlights real-world failures and use cases, advocates for context-aware explanations, and calls for a paradigm shift from identifying where models focus to understanding how they reason.
Interpretable machine learning for time-to-event prediction in medicine and healthcare
by Hubert Baniecki, Bartlomiej Sobieski, Patryk Szatkowski, Przemyslaw Bombinski, Przemyslaw Biecek
published in Artificial Intelligence in Medicine, 2025 #Surv-XAI #BioMed
This paper advances explainable survival analysis by introducing time-dependent feature effects and global feature importance methods tailored for time-to-event predictions in healthcare. Through evaluations on multi-modal and multi-omics datasets, the authors demonstrate how post-hoc interpretations can uncover model biases, guide improvements, and support clinical insights, contributing open resources for further research.
Exploring local explanations of nonlinear models using animated linear projections
by Nicholas Spyrison, Dianne Cook, Przemyslaw Biecek
published in Computational Statistics, 2025 #model-exploration #shap
This paper addresses limitations of local variable attributions (LVAs) in handling predictor interactions by introducing a method that converts LVAs into linear projections, visualized via radial tours. Implemented in the R package cheem, this approach enhances interpretability of complex models by revealing interaction effects, outliers, and prediction errors across diverse categorical and quantitative datasets.
Aggregated Attributions for Explanatory Analysis of 3D Segmentation Models
by Maciej Chrabaszcz, Hubert Baniecki, Piotr Komorowski, Szymon Plotka, Przemyslaw Biecek
published in IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025 #GenXAI #BioMed
This paper presents AGG2Exp, a novel method for aggregating voxel-level attributions to globally explain 3D medical image segmentation models, addressing limitations of traditional saliency maps in high-dimensional, multi-class contexts. Benchmarking shows AGG2Exp provides more faithful and comprehensive insights than perturbation-based methods, enabling deeper understanding of models like Swin UNETR trained on TotalSegmentator v2.
On the Robustness of Global Feature Effect Explanations
by Hubert Baniecki, Giuseppe Casalicchio, Bernd Bischl, Przemyslaw Biecek
published in European Conference on Machine Learning (ECML), 2025 #post-hoc #Theory-XAI
This paper investigates the robustness of global post-hoc explanation methods, such as partial dependence plots and accumulated local effects, in the context of tabular data. By deriving theoretical bounds and conducting empirical evaluations, the authors reveal the sensitivity of these explanations to data and model perturbations, highlighting risks of misinterpretation in applied machine learning.
Global Counterfactual Directions
by Bartlomiej Sobieski, Przemyslaw Biecek
published in European Conference on Computer Vision (ECCV), 2025 #GenXAI #Counterfactuals
This paper introduces Global Counterfactual Directions (GCDs), a novel method that extends visual counterfactual explanations from local to global analysis by uncovering classifier-relevant directions in the latent space of Diffusion Autoencoders. The approach enables black-box manipulation of classifier decisions across datasets and enhances attribution methods, offering improved interpretability and generalization in real-world applications.
Rethinking Visual Counterfactual Explanations Through Region Constraint
by Bartlomiej Sobieski, Jakub Grzywaczewski, Bartlomiej Sadlej, Matthew Tivnan, Przemyslaw Biecek
published in International Conference on Learning Representations (ICLR), 2025 #GenXAI #Counterfactuals
This paper introduces region-constrained visual counterfactual explanations (RVCEs), which limit modifications to predefined image regions to enhance interpretability and reduce cognitive biases in classifier explanations. By leveraging a novel Region-Constrained Counterfactual Schrödinger Bridge (RCSB), the approach achieves state-of-the-art performance while enabling precise and interactive counterfactual reasoning.
Efficient and Accurate Explanation Estimation with Distribution Compression
by Hubert Baniecki, Giuseppe Casalicchio, Bernd Bischl, Przemyslaw Biecek
published in International Conference on Learning Representations (ICLR) Spotlight, 2025 #post-hoc #Theory-XAI
This paper presents Compress Then Explain (CTE), a novel paradigm that enhances the efficiency and accuracy of post-hoc explanation methods by replacing i.i.d. sampling with kernel-based distribution compression. CTE reduces computational overhead and improves approximation quality of feature attributions, achieving comparable results with 2–3× fewer model evaluations across various explanation techniques.
Investigating the impact of balancing, filtering, and complexity on predictive multiplicity: A data-centric perspective
by Mustafa Cavus, Przemyslaw Biecek
published in Information Fusion, 2025 #Rashomon #AutoML
This paper explores how data preprocessing techniques, particularly balancing and filtering, influence predictive multiplicity and model stability under the Rashomon effect across 21 real-world datasets. The findings highlight the trade-offs in data-centric AI approaches, revealing that while filtering can enhance generalization, certain balancing methods may exacerbate predictive multiplicity, especially in complex datasets.
Deep spatial context: when attention-based models meet spatial regression
by Paulina Tomaszewska, Elżbieta Sienkiewicz, Mai P. Hoang, Przemyslaw Biecek
published in International Conference on Artificial Intelligence in Medicine (AIME), 2025 #Spatial-XAI #BioMed
This paper introduces DSCon, a method for analyzing attention-based vision models through quantitative spatial context measures—SCMfeatures, SCMtargets, and SCMresiduals—enabled by integrating spatial regression into the analysis pipeline. Inspired by histopathology but broadly applicable, DSCon reveals that spatial context plays a greater role in tumor classification than in normal tissues, with its influence diminishing as neighborhood size increases and being most prominent within the feature space.
Interpreting CLIP with Hierarchical Sparse Autoencoders
by Vladimir Zaigrajew, Hubert Baniecki, Przemyslaw Biecek
published in International Conference on Machine Learning (ICML), 2025 #Multimodal-XAI #SAE
This paper introduces Matryoshka SAE (MSAE), a hierarchical sparse autoencoder that optimizes both reconstruction quality and sparsity for interpreting large-scale vision-language models like CLIP. MSAE achieves a new state-of-the-art trade-off, enabling extraction of over 120 semantic concepts for tasks such as concept-based search and bias analysis, thus enhancing interpretability and control in multimodal systems.
MASCOTS: Model-Agnostic Symbolic COunterfactual explanations for Time Series
by Dawid Płudowski, Francesco Spinnato, Piotr Wilczyński, Krzysztof Kotowski, Evridiki Vasileia Ntagiou, Riccardo Guidotti, Przemyslaw Biecek
published in European Conference on Machine Learning (ECML), 2025 #TimeSeries-XAI #Counterfactuals
This paper introduces MASCOTS, a model-agnostic method for generating counterfactual explanations in time series by leveraging symbolic representations derived from Bag-of-Receptive-Fields and Symbolic Aggregate Approximation. MASCOTS enhances interpretability and sparsity while maintaining fidelity and plausibility, enabling human-accessible explanations across univariate and multivariate datasets.
talks
Talk 2 on Relevant Topic in Your Field
by
Published:
More information here
teaching
eXplainable Machine Learning
by
Graduate course, University of Warsaw, MIM, 2023
Course materials are at https://github.com/mim-uw/eXplainableMachineLearning-2024
MSc seminars, University of Warsaw, MIM, 2023
Course materials are at https://github.com/mim-uw/MachineLearningSeminar