04213ntm a22005657i 4500 000722929 CZ-PrVSE 20250608110309.0 m d cr n|||||||||| 250608s2025 xr fsbm 000 0 eng d NEZPRACOVANÝ IMPORT ABA006 cze ABA006 ABA006 rda Jeršova, Julija ISIS:178196 dis Vysvětlitelná klasifikace obrazových dat na základě slovních embeddingů eng Explainable image classification based on word embeddings / Julija Jeršova 2025 ?? stran : digital, PDF soubor Vedoucí práce: Tomáš Kliegr Diplomová práce (Ing.)—Vysoká škola ekonomická v Praze. Fakulta informatiky a statistiky, 2025 Obsahuje bibliografii Textový (vysokoškolská kvalifikační práce) Rok obhajoby 2025 As artificial intelligence systems become increasingly integrated into real-world applications, the need for transparency and user trust has never been greater. In computer vision, one of the main challenges is to explain model decisions in a way that is both accurate and understandable to non-expert users. This thesis introduces a novel method that generates visual explanations using simple mathematical reasoning, producing bounding boxes to highlight relevant areas in an image. The goal was to create a technique that would be interpretable by design and easy to evaluate. We extend the SMER (Self Model Entities Related) approach and adapt it to image data by linking word-level importance to spatial regions. The resulting bounding boxes are evaluated using a comprehensive set of strategies. To evaluate the proposed approach, we used a dataset containing a total of 13,355 images. Quantitatively, we applied the AOPC (Area Over the Perturbation Curve) metric to assess the effectiveness of explanations in preserving model performance. Qualitatively, we compare SMER-based explanations with LIME (Local Interpretable Model-Agnostic Explanations) using a logistic regression classifier, and with MoRF-based (Most Relevant First) heatmaps using a ResNet classifier to benchmark interpretability. In addition, we conducted a large-scale, multilingual human-centered evaluation. In this survey, 270 participants rated visual explanations---including bounding boxes and heatmaps---across ten categories of objects and provided open-text feedback on their understandability and clarity. The findings indicate that bounding boxes generated by our method are generally perceived as more understandable than heatmaps. While the approach shows strong potential, it faces limitations when applied to images with multiple objects or abstract concepts. Additionally, the precision of bounding boxes is constrained by the capabilities of current language models, though this is likely to improve as the models evolve. Overall, the results suggest that simple, explainable models can be powerful tools in advancing the interpretability of AI systems. Způsob přístupu: Internet kognitivní informatika [obor dipl. práce] diplomové práce fd132022 czenas master's theses eczenas word embeddings SMER LIME explainable image classification MoRF LLM bounding box heatmap Kliegr, Tomáš ISIS:8484 ths Sýkora, Lukáš ISIS:36976 opn Vysoká škola ekonomická v Praze. Fakulta informatiky a statistiky kn20010709399 dgg https://insis.vse.cz/zp/87613/podrobnosti VŠKP v InSIS https://insis.vse.cz/zp/87613 Hlavní práce https://insis.vse.cz/zp/87613/posudek/vedouci Hodnocení vedoucího https://insis.vse.cz/zp/87613/posudek/oponent/86445 Oponentura https://insis.vse.cz/zp/87613/podrobnosti dc:identifier NEPOSILAT VSKP vse87613 250605 87613