Témák listája

Computer vision (CV) and Natural Language Processing (NLP) are two main subfields of machine learning, and a lot of research is going on there. These two subfields overlap together in tasks such as text generation out of image (image2text) or vice-versa (text2image). A main challenge that is facing these models (and ML based models in general) is the explaination of the model's output, e.g.: why a certain object appears in a certain image captioning. The task of the student(s) is to get familiar with those tasks and try to reproduce available XAI (explainable AI) algorithms in order to utilize them later. Number of students: 1 - 2.

Computer vision (CV) and Natural Language Processing (NLP) are two main subfields of machine learning, and a lot of research is going on there. These two subfields overlap together in tasks such as text generation out of image (image2text) or vice-versa (text2image). A new subfield has emerged, i.e. Story Visualization, with the help of the advancement of GANs and Diffusion models. The task of the student(s) is to explore Story Visualization topic by investigating and utilizing the state-of-the-art models in the field. No. of students: 1 - 3 contact email: alshouha@edu.bme.hu

This laboratory project focuses on making AI explanations for time-series models easier for humans to understand. In previous research, we developed a fast method called Dynamic Window Perturbation Analysis to identify which time periods and variables are most important for a model’s prediction. The method produces heatmaps that highlight important temporal regions. While these results are technically faithful and computationally efficient, they are not always intuitive for end-users, such as clinicians working with medical time-series data. The goal of this project is to transform these window-level importance heatmaps into clear, human-understandable explanations. Students will design a framework that detects meaningful temporal patterns (such as increasing trends, sudden spikes, or late-stage changes) from the model’s importance scores and converts them into concise narrative statements. For example: “A sudden spike in feature X during time steps 5–7 strongly influenced the prediction.” The project combines explainable AI, time-series analysis, and natural language generation.