Institutional Repository of Vilnius University
Dideliais kalbų modeliais sugeneruotų vaizdų atpažinimas klasifikavimo uždavinyje /
Abstract
dc:descriptionNowadays, more and more tasks are being solved using artificial intelligence. Some more sensitive application areas, such as medical image analysis or autonomous driving require ensuring the reliability of the model output. In real-world applications, classifier performance often degrades due to data distribution shifts – model encounter examples that differ significantly from those seen during training and therefore often fail to make the right decisions. Such examples are referred as out-of-distribution (OOD). In classification tasks, detecting these distributional shifts is essential to ensure the safety of the models for its users. The study investigates five different out-of-distribution (OOD) detection methods – analyzing their underlying principles and evaluating their ability to detect distribution shifts using the OpenOOD benchmarking platform. The effectiveness of these OOD detection methods was assessed using standard datasets provided by the OpenOOD platform, which are designed for semantic shift detection, as well as artificially generated datasets created with the Stable Diffusion v2.0 model, which allow for the analysis of more subtle, covariate shifts. The experiment revealed that the ability of the methods to detect distribution shifts is directly proportional to the magnitude of the shift – while the methods effectively detect large semantic shifts, they perform poorly when identifying small covariate shifts. The results also showed that the PixMix method demonstrated the best performance for detecting distribution shifts.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Janavičius, Linas,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD239030474&prefLang=en_US