Översikt
- Datum:Startar 31 mars 2025, 14:00Slutar 31 mars 2025, 17:00
- Plats:Lecture Hall SB-H4, Sven Hultins Gata 6, Gothenburg
- Opponent:Professor, Fredrik Lindsten, Linköping University, Linköping, Sweden.
- AvhandlingLäs avhandlingen (Öppnas i ny flik)
Despite the success and potential of deep learning techniques, ensuring the reliable deployment of such models remains a primary concern. In this thesis, the reliability of deep models is tackled through the lens of out-of-distribution (OOD) detection, model calibration, and hallucination mitigation to contributing to a trustworthy artificial intelligence (AI) system.
Paper A and Paper B utilize joint energy-based modeling (JEM), and develop a probabilistic classifier and regressor, respectively. Specifically, Paper A addresses the training instability of joint energy-based models by replacing stochastic gradient Langevin dynamics with slice score matching, which results in a smoother training procedure without compromising the OOD performance. Paper B extends the idea of JEM from classification to regression, leading to a better calibrated regressor.
Paper C focuses on large-scale OOD detection with standard discriminative classifiers and proposes a novel OOD score based on generalized entropy, utilizing only information from the probability space.
Paper D leverages transfer learning and self-supervised learning techniques to devise an efficient framework, in which only normal samples are required for detecting anomalies in Chest X-rays.
Paper E utilizes the powerful text-image alignment in contrastive vision-language models (VLMs) for zero-shot OOD detection.
Finally, Paper F leverages insights from OOD detection and proposes an energy-based decoding method to mitigate object hallucination in generative VLMs.
Paper A and Paper B utilize joint energy-based modeling (JEM), and develop a probabilistic classifier and regressor, respectively. Specifically, Paper A addresses the training instability of joint energy-based models by replacing stochastic gradient Langevin dynamics with slice score matching, which results in a smoother training procedure without compromising the OOD performance. Paper B extends the idea of JEM from classification to regression, leading to a better calibrated regressor.
Paper C focuses on large-scale OOD detection with standard discriminative classifiers and proposes a novel OOD score based on generalized entropy, utilizing only information from the probability space.
Paper D leverages transfer learning and self-supervised learning techniques to devise an efficient framework, in which only normal samples are required for detecting anomalies in Chest X-rays.
Paper E utilizes the powerful text-image alignment in contrastive vision-language models (VLMs) for zero-shot OOD detection.
Finally, Paper F leverages insights from OOD detection and proposes an energy-based decoding method to mitigate object hallucination in generative VLMs.
