Towards Reliable Deep Foundation Models in OOD detection, model calibration, and hallucination mitigation
Doktorsavhandling, 2025
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.
Lecture Hall SB-H4, Sven Hultins Gata 6, Gothenburg
Opponent: Professor, Fredrik Lindsten, Linköping University, Linköping, Sweden.
Författare
Xixi Liu
Chalmers, Elektroteknik, Signalbehandling och medicinsk teknik
Text Prompt Augmentation for Zero-shot Out-of-Distribution Detection
2024 European Conference on Computer Vision,;Vol. 15059-15147(2024)p. 364-380
Paper i proceeding
Deep Nearest Neighbors for Anomaly Detection in Chest X-Rays
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics),;Vol. 14349 LNCS(2024)p. 293-302
Paper i proceeding
GEN: Pushing the Limits of Softmax-Based Out-of-Distribution Detection
Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition,;Vol. 2023-June(2023)p. 23946-23955
Paper i proceeding
Effortless Training of Joint Energy-Based Models with Sliced Score Matching
Proceedings - International Conference on Pattern Recognition,;Vol. 2022-August(2022)p. 2643-2649
Paper i proceeding
Joint Energy-based Model for Deep Probabilistic Regression
Proceedings - International Conference on Pattern Recognition,;Vol. 2022-August(2022)p. 2693-2699
Paper i proceeding
Energy-Guided Decoding for Object Hallucination Mitigation
Deep learning techniques have been widely utilized across various domains including autonomous driving, content generation/recommendation, drug discovery, voice assistants, and renewable energy management. Ensuring the reliability of deployed models in the real-world applications has become more critical than ever. This thesis aims to enhance the reliability of deep models for trustworthy artificial intelligence by addressing out-of distribution detection (OOD), model calibration, and hallucination mitigation.
The key results in this thesis reveal the following insights:
1) Training deep models utilizing joint energy-based modeling enhances OOD detection performance and results in better calibrated regressors and classifiers;
2) OOD detection can be effectively achieved by utilizing only information available in the probability space of discriminative classifiers;
3) Medical anomalies can be identified using only normal images. By utilizing transfer learning and self-supervised learning techniques, an efficient feature-based framework is developed to detect medical anomalies in Chest X-rays. This approach outperforms reconstruction-based methods in terms of accuracy and effectiveness;
4) The knowledge of OOD detection within the framework of discriminative classifiers, can be effectively transferred to contrastive vision-language models (VLMs), enabling zero-shot OOD detection;
5) The insight gained from OOD detection has potential to address object hallucination in generative VLMs.
The key results in this thesis reveal the following insights:
1) Training deep models utilizing joint energy-based modeling enhances OOD detection performance and results in better calibrated regressors and classifiers;
2) OOD detection can be effectively achieved by utilizing only information available in the probability space of discriminative classifiers;
3) Medical anomalies can be identified using only normal images. By utilizing transfer learning and self-supervised learning techniques, an efficient feature-based framework is developed to detect medical anomalies in Chest X-rays. This approach outperforms reconstruction-based methods in terms of accuracy and effectiveness;
4) The knowledge of OOD detection within the framework of discriminative classifiers, can be effectively transferred to contrastive vision-language models (VLMs), enabling zero-shot OOD detection;
5) The insight gained from OOD detection has potential to address object hallucination in generative VLMs.
Ämneskategorier (SSIF 2025)
Datorseende och lärande system
Robotik och automation
Signalbehandling
ISBN
978-91-8103-158-4
Doktorsavhandlingar vid Chalmers tekniska högskola. Ny serie: 5616
Utgivare
Chalmers
Lecture Hall SB-H4, Sven Hultins Gata 6, Gothenburg
Opponent: Professor, Fredrik Lindsten, Linköping University, Linköping, Sweden.