1<!DOCTYPE html> <html lang="en"> <head> <meta http-equiv="Content-Type" content="text/html; charset=UTF-8"> <meta name="google-site-verification" content="lWJ1XX1KZ5Xuffkoz13p1CgE6lp8skvbb53C8mM0Aas"> <meta http-equiv="Permissions-Policy" content="interest-cohort=()"> <meta charset="utf-8"> <meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no"> <meta http-equiv="X-UA-Compatible" content="IE=edge"> <title> Lenka TÄtková </title> <meta name="author" content="Lenka TÄtková"> <meta name="description" content="A simple, whitespace theme for academics. Based on [*folio](https://github.com/bogoli/-folio) design. "> <meta name="keywords" content="explainability, xAI"> <meta property="og:site_name" content="Lenka TÄtková"> <meta property="og:type" content="website"> <meta property="og:title" content="Lenka TÄtková | about"> <meta property="og:url" content="https://lenkatetkova.github.io/"> <meta property="og:description" content="A simple, whitespace theme for academics. Based on [*folio](https://github.com/bogoli/-folio) design. "> <meta property="og:locale" content="en"> <meta name="twitter:card" content="summary"> <meta name="twitter:title" content="about"> <meta name="twitter:description" content="A simple, whitespace theme for academics. Based on [*folio](https://github.com/bogoli/-folio) design. ">
1<script type="application/ld+json"> 2 { 3 "author": 4 { 5 "@type": "Person", 6 "name": "Lenka TÄtková" 7 }, 8 "url": "https://lenkatetkova.github.io/", 9 "@type": "WebSite", 10 "description": "A simple, whitespace theme for academics. Based on [*folio](https://github.com/bogoli/-folio) design. 11", 12 "headline": "about", 13 14 "name": "Lenka TÄtková", 15 "@context": "https://schema.org" 16 } 17 </script>
17 <link rel="stylesheet" href="/assets/css/bootstrap.min.css?a4b3f509e79c54a512b890d73235ef04"> <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/css/mdb.min.css" integrity="sha256-jpjYvU3G3N6nrrBwXJoVEYI/0zw8htfFnhT9ljN3JJw=" crossorigin="anonymous"> <link defer rel="stylesheet" href="/assets/css/academicons.min.css?f0b7046b84e425c55f3463ac249818f5"> <link defer rel="stylesheet" href="/assets/css/scholar-icons.css?62b2ac103a88034e6882a5be5f3e2772"> <link defer rel="stylesheet" type="text/css" href="https://fonts.googleapis.com/css?family=Roboto:300,400,500,700|Roboto+Slab:100,300,400,500,700|Material+Icons&display=swap"> <link defer rel="stylesheet" href="/assets/css/jekyll-pygments-themes-github.css?591dab5a4e56573bf4ef7fd332894c99" media="" id="highlight_theme_light"> <link rel="shortcut icon" href="data:image/svg+xml,<svg%20xmlns=%22http://www.w3.org/2000/svg%22%20viewBox=%220%200%20100%20100%22><text%20y=%22.9em%22%20font-size=%2290%22>%F0%9F%A4%96</text></svg>"> <link rel="stylesheet" href="/assets/css/main.css?d41d8cd98f00b204e9800998ecf8427e"> <link rel="canonical" href="https://lenkatetkova.github.io/">
17<script src="/assets/js/theme.js?a81d82887dd692e91686b43de4542f18"></script>
17 <link defer rel="stylesheet" href="/assets/css/jekyll-pygments-themes-native.css?5847e5ed4a4568527aa6cfab446049ca" media="none" id="highlight_theme_dark">
17<script> 18 initTheme(); 19 </script>
19 </head> <body class="fixed-top-nav "> <header> <nav id="navbar" class="navbar navbar-light navbar-expand-sm fixed-top" role="navigation"> <div class="container"> <button class="navbar-toggler collapsed ml-auto" type="button" data-toggle="collapse" data-target="#navbarNav" aria-controls="navbarNav" aria-expanded="false" aria-label="Toggle navigation"> <span class="sr-only">Toggle navigation</span> <span class="icon-bar top-bar"></span> <span class="icon-bar middle-bar"></span> <span class="icon-bar bottom-bar"></span> </button> <div class="collapse navbar-collapse text-right" id="navbarNav"> <ul class="navbar-nav ml-auto flex-nowrap"> <li class="nav-item active"> <a class="nav-link" href="/">about <span class="sr-only">(current)</span> </a> </li> <li class="nav-item "> <a class="nav-link" href="/news/">news </a> </li> <li class="nav-item "> <a class="nav-link" href="/publications/">publications </a> </li> <li class="nav-item "> <a class="nav-link" href="/repositories/">repositories </a> </li> <li class="nav-item "> <a class="nav-link" href="/teaching/">teaching </a> </li> <li class="nav-item"> <button id="search-toggle" title="Search" onclick="openSearchModal()"> <span class="nav-link">ctrl k <i class="ti ti-search"></i></span> </button> </li> <li class="toggle-container"> <button id="light-toggle" title="Change theme"> <i class="ti ti-sun-moon" id="light-toggle-system"></i> <i class="ti ti-moon-filled" id="light-toggle-dark"></i> <i class="ti ti-sun-filled" id="light-toggle-light"></i> </button> </li> </ul> </div> </div> </nav> <progress id="progress" value="0"> <div class="progress-container"> <span class="progress-bar"></span> </div> </progress> </header> <div class="container mt-5" role="main"> <div class="post"> <header class="post-header"> <h1 class="post-title"> <span class="font-weight-bold">Lenka</span> TÄtková </h1> <p class="desc">DDSA Postdoctoral Fellow · Explainable AI</p> </header> <article> <div class="profile float-right"> <figure> <picture> <source class="responsive-img-srcset" srcset="/assets/img/IMG_7341-min-480.webp 480w,/assets/img/IMG_7341-min-800.webp 800w,/assets/img/IMG_7341-min-1400.webp 1400w," type="image/webp" sizes="(min-width: 930px) 270.0px, (min-width: 576px) 30vw, 95vw"> <img src="/assets/img/IMG_7341-min.jpg?78f7768d8045196753b23eb5ced2fba3" class="img-fluid z-depth-1 rounded" width="100%" height="auto" alt="IMG_7341-min.jpg" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> </source></picture> </figure> <div class="more-info"> <p>Richard Petersens Plads, Building 321</p> <p>Kongens Lyngby, 2800, Denmark</p> </div> </div> <div class="clearfix"> <p>I am a postdoctoral researcher at the <a href="https://cogsys.compute.dtu.dk/" rel="external nofollow noopener" target="_blank">Section for Cognitive Systems</a> (<a href="https://www.dtu.dk/" rel="external nofollow noopener" target="_blank">Technical university of Denmark</a>), affiliated with <a href="https://www.copenlu.com/" rel="external nofollow noopener" target="_blank">CopeNLU</a> (<a href="https://www.ku.dk/" rel="external nofollow noopener" target="_blank">University of Copenhagen</a>) and the <a href="https://www.aicentre.dk/" rel="external nofollow noopener" target="_blank">Pioneer Centre for AI</a>. My research interests include concept-based explainability, exploring representations in hidden layers, human-machine alignment, and translating theory from cognitive science into the context of machine learning.</p> <p>From September 2026, I am supported by a <a href="https://ddsa.dk/" rel="external nofollow noopener" target="_blank">DDSA</a> Postdoctoral Fellowship for my project <em>Geometry of Trust: A Unified Framework for Convex Latent Steering and Causal Concept Alignment</em>. The project asks whether AI safety can come from inside the model â from the geometric structure of how it represents information â rather than only from filtering inputs and outputs. It is carried out in collaboration with <a href="https://www2.compute.dtu.dk/~gear/" rel="external nofollow noopener" target="_blank">Georgios Arvanitidis</a>, <a href="https://apepa.github.io/" rel="external nofollow noopener" target="_blank">Pepa Atanasova</a>, <a href="https://www2.compute.dtu.dk/~afhar/" rel="external nofollow noopener" target="_blank">Aasa Feragen</a> and <a href="https://wickstrom.github.io/" rel="external nofollow noopener" target="_blank">Kristoffer Wickstrøm</a>.</p> <p>I have a PhD from DTU Compute, under the supervision of Professor Lars Kai Hansen. The focus of my PhD was enhancing and explaining AI with a focus on biological data.</p> </div> <h2> <a href="/news/" style="color: inherit">news</a> </h2> <div class="news"> <div class="table-responsive" style="max-height: 60vw"> <table class="table table-sm table-borderless"> <tr> <th scope="row" style="width: 20%">Jul 25, 2026</th> <td> I am giving an invited talk at the <a href="https://aiverification.org/2026/" rel="external nofollow noopener" target="_blank"><strong>9th International Symposium on AI Verification (SAIV 2026)</strong></a>, co-located with the Federated Logic Conference (FLoC) in Lisbon on July 24â25, 2026. My talk <strong>Latent space navigation â interpretation, probing and steering</strong> builds on the <a href="https://d3aconference.dk/latent-space-navigation-interpretation-probing-and-steering/" rel="external nofollow noopener" target="_blank">parallel session</a> I co-organised at the D3A conference last year. The <a href="/assets/pdf/slides_SAIV_2026.pdf"><strong>
19slides are available here</strong></a>, and hands-on notebooks are on <a href="https://github.com/LenkaTetkova/Latent-space-navigation" rel="external nofollow noopener" target="_blank">GitHub</a>. </td> </tr> <tr> <th scope="row" style="width: 20%">Jun 26, 2026</th> <td> I am honoured to have been awarded a <strong><a href="https://ddsa.dk/" rel="external nofollow noopener" target="_blank">DDSA</a> Postdoctoral Fellowship</strong> for my project <em>Geometry of Trust: A Unified Framework for Convex Latent Steering and Causal Concept Alignment</em>, starting September 2026, based at the <a href="https://cogsys.compute.dtu.dk/" rel="external nofollow noopener" target="_blank">Section for Cognitive Systems</a> at DTU Compute. The project asks whether AI safety can come from inside the model â from the geometric structure of how it represents information â rather than only from filtering inputs and outputs, building on our <a href="https://www.nature.com/articles/s41467-025-60809-y" rel="external nofollow noopener" target="_blank">Nature Communications paper</a>. </td> </tr> <tr> <th scope="row" style="width: 20%">Apr 27, 2026</th> <td> I gave a talk at <a href="https://www.fossanalytics.com/" rel="external nofollow noopener" target="_blank">FOSS Analytical A/S</a> about explainability in deep learning â covering our paper <a href="https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0333965" rel="external nofollow noopener" target="_blank"><strong>Challenges in explaining deep learning models for data with biological variation</strong></a>, a result of my PhD collaboration with FOSS, and our work on <a href="https://www.nature.com/articles/s41467-025-60809-y" rel="external nofollow noopener" target="_blank"><strong>convex decision regions in deep network representations</strong></a> published in Nature Communications. </td> </tr> <tr> <th scope="row" style="width: 20%">Feb 04, 2026</th> <td> Our paper <a href="https://ieeexplore.ieee.org/document/11460589" rel="external nofollow noopener" target="_blank"><strong>Large Vision Models Can Solve Mental Rotation Problems</strong></a> was accepted to ICASSP 2026. Take a look at our more approachable <a href="https://medium.com/@sebastianraymason_69102/can-large-vision-models-solve-mental-rotation-tasks-8b78ec7271ed" rel="external nofollow noopener" target="_blank">Medium post</a> about the paper. It is also selected for an <strong>oral presentation</strong> at the <a href="https://cmsworkshops.com/ICASSP2026/view_session.php?SessionID=1047" rel="external nofollow noopener" target="_blank">MLSP-L19: Multimodal and Contrastive Representation Learning</a> session. </td> </tr> <tr> <th scope="row" style="width: 20%">Jan 15, 2026</th> <td> I am co-organizing the ð±ððµ <a href="https://xai4cv-workshop.github.io/xai4cv2026/" rel="external nofollow noopener" target="_blank">ð«ððð°ðð© ðªð¼ð¿ð¸ððµð¼ð½</a> at CVPR 2026. </td> </tr> </table> </div> </div> <h2> <a href="/publications/" style="color: inherit">selected publications</a> </h2> <div class="publications"> <ol class="bibliography"> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">Nat Commun</abbr> <figure> <picture> <img src="/assets/img/publication_preview/convexity.png" class="preview z-depth-1 rounded" width="100%" height="auto" alt="convexity.png" data-zoomable loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> </picture> </figure> </div> <div id="tetkova2025convexity" class="col-sm-8"> <div class="title">On convex decision regions in deep network representations</div> <div class="author"> <em>Lenka TÄtková</em>, Thea Brüsch, Teresa Dorszewski, Fabian Martin Mager, Rasmus Ãrtoft Aagaard, Jonathan Foldager, Tommy Sonne Alstrøm, and Lars Kai Hansen </div> <div class="periodical"> <em>Nature Communications</em>. Also presented at ICLR 2024 Workshop on Representational Alignment (Re-Align) , Jul 2025 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="https://doi.org/10.1038/s41467-025-60809-y" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">DOI</a> <a href="http://arxiv.org/abs/2305.17154" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">arXiv</a> <a href="https://youtu.be/qXBT_-w3Mfs?feature=shared&t=1795" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> <a href="https://github.com/LenkaTetkova/Convexity-of-representations" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> <a href="/assets/pdf/poster_convexity.pdf" class="btn btn-sm z-depth-0" role="button">Poster</a> </div> <div class="abstract hidden"> <p>Current work on human-machine alignment aims at understanding machine-learned latent spaces and their relations to human representations. We study the convexity of concept regions in machine-learned latent spaces, inspired by Gärdenforsâ conceptual spaces. In cognitive science, convexity is found to support generalization, few-shot learning, and interpersonal alignment. We develop tools to measure convexity in sampled data and evaluate it across layers of state-of-the-art deep networks. We show that convexity is robust to relevant latent space transformations and, hence, meaningful as a quality of machine-learned latent spaces. We find pervasive approximate convexity across domains, including image, text, audio, human activity, and medical data. Fine-tuning generally increases convexity, and the level of convexity of class label regions in pretrained models predicts subsequent fine-tuning performance. Our framework allows investigation of layered latent representations and offers new insights into learning mechanisms, human-machine alignment, and potential improvements in model generalization.</p> </div> </div> </div> </li> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">ICASSP 2026</abbr> <figure> <picture> <img src="/assets/img/publication_preview/mental_rotation.png" class="preview z-depth-1 rounded" width="100%" height="auto" alt="mental_rotation.png" data-zoomable loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> </picture> </figure> </div> <div id="mason2025large" class="col-sm-8"> <div class="title">Large Vision Models Can Solve Mental Rotation Problems</div> <div class="author"> Sebastian Ray Mason, Anders Gjølbye, Phillip Chavarria Højbjerg, <em>Lenka TÄtková</em>, and Lars Kai Hansen </div> <div class="periodical"> <em>ICASSP 2026</em>, Sep 2025 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a>
19 <a href="https://doi.org/10.1109/icassp55912.2026.11460589" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">DOI</a> <a href="http://arxiv.org/abs/2509.15271" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">arXiv</a> <a href="https://github.com/gjoelbye/vit-mental-rotation" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> </div> <div class="abstract hidden"> <p>Mental rotation is a key test of spatial reasoning in humans and has been central to understanding how perception supports cognition. Despite the success of modern vision transformers, it is still unclear how well these models develop similar abilities. In this work, we present a systematic evaluation of ViT, CLIP, DINOv2, and DINOv3 across a range of mental-rotation tasks, from simple block structures similar to those used by Shepard and Metzler to study human cognition, to more complex block figures, three types of text, and photo-realistic objects. By probing model representations layer by layer, we examine where and how these networks succeed. We find that i) self-supervised ViTs capture geometric structure better than supervised ViTs; ii) intermediate layers perform better than final layers; iii) task difficulty increases with rotation complexity and occlusion, mirroring human reaction times and suggesting similar constraints in embedding space representations.</p> </div> </div> </div> </li> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">xAI 2025</abbr> <figure> <picture> <img src="/assets/img/publication_preview/colors_to_concepts.png" class="preview z-depth-1 rounded" width="100%" height="auto" alt="colors_to_concepts.png" data-zoomable loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> </picture> </figure> </div> <div id="dorszewski2025colors" class="col-sm-8"> <div class="title">From Colors to Classes: Emergence of Concepts in Vision Transformers</div> <div class="author"> Teresa Dorszewski, <em>Lenka TÄtková</em>, Robert Jenssen, Lars Kai Hansen, and Kristoffer Knutsen Wickstrøm </div> <div class="periodical"> <em>In World Conference on Explainable Artificial Intelligence</em>. Won the <strong>best poster award</strong> at Danish Digitalization, Data Science and AI 3.0 (D3A 2025). , Jul 2025 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="https://doi.org/10.1007/978-3-032-08317-3_2" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">DOI</a> <a href="http://arxiv.org/abs/2503.24071" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">arXiv</a> <a href="https://youtu.be/pSPuvpABKbM?si=epECim1EI-uc4pGI" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> <a href="https://github.com/teresa-sc/concepts_in_ViTs" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> <a href="/assets/pdf/poster_ColorsToClasses.pdf" class="btn btn-sm z-depth-0" role="button">Poster</a> </div> <div class="abstract hidden"> <p>Vision Transformers (ViTs) are increasingly utilized in various computer vision tasks due to their powerful representation capabilities. However, it remains understudied how ViTs process information layer by layer. Numerous studies have shown that convolutional neural networks (CNNs) extract features of increasing complexity throughout their layers, which is crucial for tasks like domain adaptation and transfer learning. ViTs, lacking the same inductive biases as CNNs, can potentially learn global dependencies from the first layers due to their attention mechanisms. Given the increasing importance of ViTs in computer vision, there is a need to improve the layer-wise understanding of ViTs. In this work, we present a novel, layer-wise analysis of concepts encoded in state-of-the-art ViTs using neuron labeling. Our findings reveal that ViTs encode concepts with increasing complexity throughout the network. Early layers primarily encode basic features such as colors and textures, while later layers represent more specific classes, including objects and animals. As the complexity of encoded concepts increases, the number of concepts represented in each layer also rises, reflecting a more diverse and specific set of features. Additionally, different pretraining strategies influence the quantity and category of encoded concepts, with finetuning to specific downstream tasks generally reducing the number of encoded concepts and shifting the concepts to more relevant categories.</p> </div> </div> </div> </li> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">ICASSP 2025</abbr> <figure> <picture> <img src="/assets/img/publication_preview/redundancy.png" class="preview z-depth-1 rounded" width="100%" height="auto" alt="redundancy.png" data-zoomable loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> </picture> </figure> </div> <div id="dorszewski2025redundant" class="col-sm-8"> <div class="title">
19How Redundant Is the Transformer Stack in Speech Representation Models?</div> <div class="author"> Teresa Dorszewski, Albert Kjøller Jacobsen, <em>Lenka TÄtková</em>, and Lars Kai Hansen </div> <div class="periodical"> <em>In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)</em>. Also presented at 4th NeurIPS Efficient Natural Language and Speech Processing Workshop (ENLSP-IV 2024) â <strong>Runner up for the best short paper award at the workshop</strong> , Apr 2025 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="https://doi.org/10.1109/ICASSP49660.2025.10887866" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">DOI</a> <a href="http://arxiv.org/abs/2409.16302" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">arXiv</a> <a href="https://lenkatetkova.github.io/assets/video/ICASSP_presentation_5825.mp4" class="btn btn-sm z-depth-0" role="button">Video</a> <a href="/assets/pdf/poster_redundancy.pdf" class="btn btn-sm z-depth-0" role="button">Poster</a> </div> <div class="abstract hidden"> <p>Self-supervised speech representation models, particularly those leveraging transformer architectures, have demonstrated remarkable performance across various tasks such as speech recognition, speaker identification, and emotion detection. Recent studies on transformer models revealed high redundancy between layers and the potential for significant pruning, which we will investigate here for transformer-based speech representation models. We perform a detailed analysis of layer similarity in speech representation models using three similarity metrics: cosine similarity, centered kernel alignment, and mutual nearest-neighbor alignment. Our findings reveal a block-like structure of high similarity, suggesting two main processing steps and significant redundancy of layers. We demonstrate the effectiveness of pruning transformer-based speech representation models without the need for post-training, achieving up to 40% reduction in transformer layers while maintaining over 95% of the modelâs predictive capacity. Furthermore, we employ a knowledge distillation method to substitute the entire transformer stack with mimicking layers, reducing the network size by 95-98% and the inference time by up to 94%. This substantial decrease in computational load occurs without considerable performance loss, suggesting that the transformer stack is almost completely redundant for downstream applications of speech representation models.</p> </div> </div> </div> </li> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">xAI 2024</abbr> <figure> <picture> <img src="/assets/img/publication_preview/knowledge_graphs.png" class="preview z-depth-1 rounded" width="100%" height="auto" alt="knowledge_graphs.png" data-zoomable loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> </picture> </figure> </div> <div id="tetkova2024knowledge" class="col-sm-8"> <div class="title">Knowledge graphs for empirical concept retrieval</div> <div class="author"> <em>Lenka TÄtková</em>, Teresa Karen Scheidt, Maria Mandrup Fogh, Ellen Marie Gaunby Jørgensen, Finn Ã
rup Nielsen, and Lars Kai Hansen </div> <div class="periodical"> <em>In World Conference on Explainable Artificial Intelligence</em>, Jul 2024 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="https://doi.org/10.1007/978-3-031-63787-2_9" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">DOI</a> <a href="http://arxiv.org/abs/2404.07008" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">arXiv</a> <a href="https://github.com/LenkaTetkova/Knowledge-Graphs-for-Empirical-Concept-Retrieval" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> </div> <div class="abstract hidden"> <p>Concept-based explainable AI is promising as a tool to improve the understanding of complex models at the premises of a given user, viz. as a tool for personalized explainability. An important class of concept-based explainability methods is constructed with empirically defined concepts, indirectly defined through a set of positive and negative examples, as in the TCAV approach (Kim et al., 2018). While it is appealing to the user to avoid formal definitions of concepts and their operationalization, it can be challenging to establish relevant concept datasets. Here, we address this challenge using general knowledge graphs (such as, e.g., Wikidata or WordNet) for c
19omprehensive concept definition and present a workflow for user-driven data collection in both text and image domains. The concepts derived from knowledge graphs are defined interactively, providing an opportunity for personalization and ensuring that the concepts reflect the userâs intentions. We test the retrieved concept datasets on two concept-based explainability methods, namely concept activation vectors (CAVs) and concept activation regions (CARs) (Crabbe and van der Schaar, 2022). We show that CAVs and CARs based on these empirical concept datasets provide robust and accurate explanations. Importantly, we also find good alignment between the modelsâ representations of concepts and the structure of knowledge graphs, i.e., human representations. This supports our conclusion that knowledge graph-based concepts are relevant for XAI.</p> </div> </div> </div> </li> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">XAI4CV @ CVPR 2023</abbr> <figure> <picture> <img src="/assets/img/publication_preview/robustness.png" class="preview z-depth-1 rounded" width="100%" height="auto" alt="robustness.png" data-zoomable loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> </picture> </figure> </div> <div id="tetkova2023robustness" class="col-sm-8"> <div class="title">Robustness of Visual Explanations to Common Data Augmentation Methods</div> <div class="author"> <em>Lenka TÄtková</em>, and Lars Kai Hansen </div> <div class="periodical"> <em>Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition</em>, Jun 2023 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="https://doi.org/10.1109/CVPRW59228.2023.00381" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">DOI</a> <a href="http://arxiv.org/abs/2304.08984" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">arXiv</a> <a href="https://github.com/LenkaTetkova/robustness-of-explanations" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> <a href="/assets/pdf/poster_robustness.pdf" class="btn btn-sm z-depth-0" role="button">Poster</a> </div> <div class="abstract hidden"> <p>As the use of deep neural networks continues to grow, understanding their behaviour has become more crucial than ever. Post-hoc explainability methods are a potential solution, but their reliability is being called into question. Our research investigates the response of post-hoc visual explanations to naturally occurring transformations, often referred to as augmentations. We anticipate explanations to be invariant under certain transformations, such as changes to the colour map while responding in an equivariant manner to transformations like translation, object scaling, and rotation. We have found remarkable differences in robustness depending on the type of transformation, with some explainability methods (such as LRP composites and Guided Backprop) being more stable than others. We also explore the role of training with data augmentation. We provide evidence that explanations are typically less robust to augmentation than classification performance, regardless of whether data augmentation is used in training or not.</p> </div> </div> </div> </li> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">Re-Align @ ICLR 2025</abbr> <figure> <picture> <img src="/assets/img/publication_preview/SLM.png" class="preview z-depth-1 rounded" width="100%" height="auto" alt="SLM.png" data-zoomable loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> </picture> </figure> </div> <div id="linhardt2025cat" class="col-sm-8"> <div class="title">Cat, Rat, Meow: On the Alignment of Language Model and Human Term-Similarity Judgments</div> <div class="author"> Lorenz Linhardt, Tom Neuhäuser, <em>Lenka TÄtková</em>, and Oliver Eberle </div> <div class="periodical"> <em>In Second Workshop on Representational Alignment at ICLR 2025</em>, Apr 2025 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="https://openreview.net/pdf?id=wIcg6M2Dvj" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> </div> <div class="abstract hidden"> <p>Small and mid-sized generative language models have gained increasing attention. Their size and availability make them amenable to be analyzed at a behavioral as well as a representational level, allowing investigations of how these levels interact. We evaluate 32 publicly available language models for their representational and behavioral alignment with human similarity judgments on a word triplet task. This provides a novel evaluation setting to probe semantic associations in language beyond common pairwise comparisons. We find that (1) even the representations of small language models can achieve human-level alignment, (2) instruction-tuned model variants can exhibit substantially increased agreement, (3) the pattern of alignment across layers is highly model dependent, and (4) alignment based on modelsâ behavioral responses is highly depen
19dent on model size, matching their representational alignment only for the largest evaluated models.</p> </div> </div> </div> </li> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">NLDL 2025</abbr> <figure> <picture> <img src="/assets/img/publication_preview/convexity_vs_oooa.png" class="preview z-depth-1 rounded" width="100%" height="auto" alt="convexity_vs_oooa.png" data-zoomable loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> </picture> </figure> </div> <div id="dorszewski2024connecting" class="col-sm-8"> <div class="title">Connecting Concept Convexity and Human-Machine Alignment in Deep Neural Networks</div> <div class="author"> Teresa Dorszewski, <em>Lenka TÄtková</em>, Lorenz Linhardt, and Lars Kai Hansen </div> <div class="periodical"> <em>Proceedings of the 6th Northern Lights Deep Learning Conference (NLDL)</em>, Jan 2025 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="http://arxiv.org/abs/2409.06362" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">arXiv</a> <a href="https://raw.githubusercontent.com/mlresearch/v265/main/assets/dorszewski25a/dorszewski25a.pdf" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> <a href="/assets/pdf/poster_nldl.pdf" class="btn btn-sm z-depth-0" role="button">Poster</a> </div> <div class="abstract hidden"> <p>Understanding how neural networks align with human cognitive processes is a crucial step toward developing more interpretable and reliable AI systems. Motivated by theories of human cognition, this study examines the relationship between convexity in neural network representations and human-machine alignment based on behavioral data. We identify a correlation between these two dimensions in pretrained and fine-tuned vision transformer models. Our findings suggest that the convex regions formed in latent spaces of neural networks to some extent align with human-defined categories and reflect the similarity relations humans use in cognitive tasks. While optimizing for alignment generally enhances convexity, increasing convexity through fine-tuning yields inconsistent effects on alignment, which suggests a complex relationship between the two. This study presents a first step toward understanding the relationship between the convexity of latent representations and human-machine alignment.</p> </div> </div> </div> </li> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">IEEE MLSP 2024</abbr> <figure> <picture> <img src="/assets/img/publication_preview/convexity_pruning.png" class="preview z-depth-1 rounded" width="100%" height="auto" alt="convexity_pruning.png" data-zoomable loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> </picture> </figure> </div> <div id="dorszewski2024convexity" class="col-sm-8"> <div class="title">Convexity Based Pruning of Speech Representation Models</div> <div class="author"> Teresa Dorszewski, <em>Lenka TÄtková</em>, and Lars Kai Hansen </div> <div class="periodical"> <em>In 2024 IEEE 34th International Workshop on Machine Learning for Signal Processing (MLSP)</em>, Sep 2024 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="https://doi.org/10.1109/MLSP58920.2024.10734716" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">DOI</a> <a href="http://arxiv.org/abs/2408.11858" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">arXiv</a> <a href="https://youtu.be/pSPuvpABKbM?si=epECim1EI-uc4pGI" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> <a href="/assets/pdf/poster_mlsp.pdf" class="btn btn-sm z-depth-0" role="button">Poster</a> </div> <div class="abstract hidden"> <p>Speech representation models based on the transformer architecture and trained by self-supervised learning have shown great promise for solving tasks such as speech and speaker recognition, keyword spotting, emotion detection, and more. Typically, it is found that larger models lead to better performance. However, the significant computational effort involved in such large transformer systems is a challenge for embedded and real-world applications. Recent work has shown that there is significant redundancy in the transformer models for NLP and massive layer pruning is feasible (Sajjad et al., 2023). Here, we investigate layer pruning in audio models. We base the pruning decision on a convexity criterion. Convexity of classification regions has recently been proposed as an indicator of subsequent fine-tuning performance in a range of application domains, including NLP and audio. In empirical investigations, we find a massive reduction in the computational effort with no loss of performance or even improvements in certain cases</p> </div> </div> </div> </li> <li> <div class="row"> <div class="col col-sm-2 abbr"> <abbr class="badge rounded w-100">PloS one</abbr> <figure> <picture> <img src="/assets/img/publication_preview/challenges.jpg" class="preview z-depth-1 rounded" width="100%" height="auto" alt="challenges.jpg" data-zoomable loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> </picture> </figure> </div> <div id="tetkova2025challenges" class="col-sm-8"> <div class="title">Challenges in explaining deep learning models for data with biological variation</div> <div class="author"> <em>Lenka TÄtková</em>, Erik Schou Dreier, Robin Malm, and Lars Kai Hansen </div> <div class="periodical"> <em>PloS one</em>, Oct 2025 </div> <div class="periodical"> </div> <div class="links"> <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> <a href="https://doi.org/10.1371/journal.pone.0333965" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">DOI</a> <a href="http://arxiv.org/abs/2406.09981" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">arXiv</a> <a href="https://github.com/LenkaTetkova/Challenges-in-explaining-deep-learning-models-for-data-with-biological-variation.git" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> <a href="/assets/pdf/poster_grains.pdf" class="btn btn-sm z-depth-0" role="button">Poster</a> </div> <div class="abstract hidden"> <p>Much machine learning research progress is based on developing models and evaluating them on a benchmark dataset (e.g., ImageNet for images). However, applying such benchmark-successful methods to real-world data often does not work as expected. This is particularly the case for biological data where we expect variability at multiple time and spatial scales. Typical benchmark data has simple, dominant semantics, such as a number, an object type, or a word. In contrast, biological samples often have multiple semantic components leading to complex and entangled signals. Complexity is added if the signal of interest is related to atypical states, e.g., disease, and if there is limited data available for learning. In this work, we focus on image classification of real-world biological data that are, indeed, different from standard images. We are using grain data and the goal is to detect diseases and damages, for example, âpink fusarium" and âskinned". Pink fusarium, skinned grains, and other diseases and damages are key factors in setting the price of grains or excluding dangerous grains from food production. Apart from challenges stemming from differences of the data from the standard toy datasets, we also present challenges that need to be overcome when explaining deep learning models. For example, explainability methods have many hyperparameters that can give different results, and the ones published in the papers do not work on dissimilar images. Other
19challenges are more general: problems with visualization of the explanations and their comparison since the magnitudes of their values differ from method to method. An open fundamental question also is: How to evaluate explanations? It is a non-trivial task because the âground truth" is usually missing or ill-defined. Also, human annotators may create what they think is an explanation of the task at hand, yet the machine learning model might solve it in a different and perhaps counter-intuitive way. We discuss several of these challenges and evaluate various post-hoc explainability methods on grain data. We focus on robustness, quality of explanations, and similarity to particular âground truth" annotations made by experts. The goal is to find the methods that overall perform well and could be used in this challenging task. We hope that the proposed pipeline would be used as a framework for evaluating explainability methods in specific use cases.</p> </div> </div> </div> </li> </ol> </div> <div class="social"> <div class="contact-icons"> <a href="https://bsky.app/profile/lenkatetkova.bsky.social" title="Bluesky" rel="external nofollow noopener" target="_blank"><i class="fa-brands fa-bluesky"></i></a> <a href="https://dblp.org/pid/345/1472.html" title="DBLP" rel="external nofollow noopener" target="_blank"><i class="ai ai-dblp"></i></a> <a href="mailto:%6C%65%6E%68%79@%64%74%75.%64%6B" title="email"><i class="fa-solid fa-envelope"></i></a> <a href="https://github.com/lenkatetkova" title="GitHub" rel="external nofollow noopener" target="_blank"><i class="fa-brands fa-github"></i></a> <a href="https://www.linkedin.com/in/lenka-tetkova" title="LinkedIn" rel="external nofollow noopener" target="_blank"><i class="fa-brands fa-linkedin"></i></a> <a href="https://orcid.org/0000-0002-0009-6896" title="ORCID" rel="external nofollow noopener" target="_blank"><i class="ai ai-orcid"></i></a> <a href="https://scholar.google.com/citations?user=GSJaf6kAAAAJ" title="Google Scholar" rel="external nofollow noopener" target="_blank"><i class="ai ai-google-scholar"></i></a> </div> <div class="contact-note">You can even add a little note about which of these is the best way to reach you. </div> </div> </article> </div> </div> <footer class="fixed-bottom" role="contentinfo"> <div class="container mt-0"> © Copyright 2026 Lenka TÄtková. Powered by <a href="https://jekyllrb.com/" target="_blank" rel="external nofollow noopener">Jekyll</a> with <a href="https://github.com/alshedivat/al-folio" rel="external nofollow noopener" target="_blank">al-folio</a> theme. Hosted by <a href="https://pages.github.com/" target="_blank" rel="external nofollow noopener">GitHub Pages</a>. </div> </footer>
19<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/jquery.min.js" integrity="sha256-/xUj+3OJU5yExlq6GSYGSHk7tPXikynS7ogEvDej/m4=" crossorigin="anonymous"></script>
19
19<script src="/assets/js/bootstrap.bundle.min.js"></script>
19
19<script src="https://cdn.jsdelivr.net/npm/[email protected]/js/mdb.min.js" integrity="sha256-NdbiivsvWt7VYCt6hYNT3h/th9vSTL4EDWeGs5SN3DA=" crossorigin="anonymous"></script>
19
19<script defer src="https://cdn.jsdelivr.net/npm/[email protected]/dist/masonry.pkgd.min.js" integrity="sha256-Nn1q/fx0H7SNLZMQ5Hw5JLaTRZp0yILA/FRexe19VdI=" crossorigin="anonymous"></script>
19
19<script defer src="https://cdn.jsdelivr.net/npm/[email protected]/imagesloaded.pkgd.min.js" integrity="sha256-htrLFfZJ6v5udOG+3kNLINIKh2gvoKqwEhHYfTTMICc=" crossorigin="anonymous"></script>
19
19<script defer src="/assets/js/masonry.js?a0db7e5d5c70cc3252b3138b0c91dcaf" type="text/javascript"></script>
19
19<script defer src="https://cdn.jsdelivr.net/npm/[email protected]/dist/medium-zoom.min.js" integrity="sha256-ZgMyDAIYDYGxbcpJcfUnYwNevG/xi9OHKaR/8GK+jWc=" crossorigin="anonymous"></script>
19
19<script defer src="/assets/js/zoom.js?85ddb88934d28b74e78031fd54cf8308"></script>
19
19<script src="/assets/js/no_defer.js?2781658a0a2b13ed609542042a859126"></script>
19
19<script defer src="/assets/js/common.js?e0514a05c5c95ac1a93a8dfd5249b92e"></script>
19
19<script defer src="/assets/js/copy_code.js?c8a01c11a92744d44b093fc3bda915df" type="text/javascript"></script>
19
19<script defer src="/assets/js/jupyter_new_tab.js?d9f17b6adc2311cbabd747f4538bb15f"></script>
19
19<script async src="https://d1bxh8uas1mnw7.cloudfront.net/assets/embed.js"></script>
19
19<script async src="https://badge.dimensions.ai/badge.js"></script>
19
19<script defer type="text/javascript" id="MathJax-script" src="https://cdn.jsdelivr.net/npm/[email protected]/es5/tex-mml-chtml.js" integrity="sha256-MASABpB4tYktI2Oitl4t+78w/lyA+D7b/s9GEP0JOGI=" crossorigin="anonymous"></script>
19
19<script src="/assets/js/mathjax-setup.js?a5bb4e6a542c546dd929b24b8b236dfd"></script>
19
19<script defer src="https://cdnjs.cloudflare.com/polyfill/v3/polyfill.min.js?features=es6" crossorigin="anonymous"></script>
vendor: 64 bytes, line 19
19 <script async src="https://www.googletagmanager.com/gtag/js?id=
19G-FBZ0CXEEHD
vendor: 12 bytes, line 19
19"></script>
19<script> 20
vendor: 153 bytes, lines 20-26
20window.dataLayer = window.dataLayer || []; 21 function gtag() { 22 dataLayer.push(arguments); 23 } 24 gtag('js', new Date()); 25 26 gtag('config', '
26G-FBZ0CXEEHD
vendor: 6 bytes, lines 26-27
26'); 27
27</script>
27
27<script defer src="/assets/js/google-analytics-setup.js?12374742c4b1801ba82226e617af7e2d"></script>
27
27<script defer src="/assets/js/progress-bar.js?2f30e0e6801ea8f5036fa66e1ab0a71a" type="text/javascript"></script>
27
27<script src="/assets/js/vanilla-back-to-top.min.js?f40d453793ff4f64e238e420181a1d17"></script>
27
27<script> 28 addBackToTop(); 29 </script>
29
29<script type="module" src="/assets/js/search/ninja-keys.min.js?a3446f084dcaecc5f75aa1757d087dcf"></script>
29 <ninja-keys hidebreadcrumbs noautoloadmdicons placeholder="Type to start searching"></ninja-keys>
29<script src="/assets/js/search-setup.js?6c304f7b1992d4b60f7a07956e52f04a"></script>
29
29<script src="/assets/js/search-data.js"></script>
29
29<script src="/assets/js/shortcut-key.js?6f508d74becd347268a7f822bca7309d"></script>
29 </body> </html>
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.