Exploring probabilistic models for semi-supervised learning

Thesis

Deep neural networks are increasingly harnessed for computer vision tasks, thanks to their robust performance. However, their training demands large-scale labeled datasets, which are labor-intensive to prepare. Semi-supervised learning (SSL) offers a solution by learning from a mix of labeled and unlabeled data.

While most state-of-the-art SSL methods follow a deterministic approach, the exploration of their probabilistic counterparts remains limited. This research area is important because probabilistic models can provide uncertainty estimates critical for real-world applications. For instance, SSL-trained models may fall short of those trained with supervised learning due to potential pseudo-label errors in unlabeled data, and these models are more likely to make wrong predictions in practice. Especially in critical sectors like medical image analysis and autonomous driving, decision-makers must understand the model’s limitations and when incorrect predictions may occur, insights often provided by uncertainty estimates. Furthermore, uncertainty can also serve as a criterion for filtering out unreliable pseudo-labels when unlabeled samples are used for training, potentially improving deep model performance.

This thesis furthers the exploration of probabilistic models for SSL. Drawing on the widely-used Bayesian approximation tool, Monte Carlo (MC) dropout, I propose a new probabilistic framework, the Generative Bayesian Deep Learning (GBDL) architecture, for semi-supervised medical image segmentation. This approach not only mitigates potential overfitting found in previous methods but also achieves superior results across four evaluation metrics. Unlike its empirically designed predecessors, GBDL is underpinned by a full Bayesian formulation, providing a theoretical probabilistic foundation.

Acknowledging MC dropout’s limitations, I introduce NP-Match, a novel proba- bilistic approach for large-scale semi-supervised image classification. We evaluated NP-Match’s generalization capabilities through extensive experiments in different challenging settings such as standard, imbalanced, and multi-label semi-supervised image classification. According to the experimental results, NP-Match not only competes favorably with previous state-of-the-art methods but also estimates uncertainty more rapidly than MC-dropout-based models, thus enhancing both training and testing efficiency.

Lastly, I propose NP-SemiSeg, a new probabilistic model for semi-supervised se- mantic segmentation. This flexible model can be integrated with various existing segmentation frameworks to make predictions and estimate uncertainty. Experiments indicate that NP-SemiSeg surpasses MC dropout in accuracy, uncertainty quantification, and speed.

Authors: Wang, J

• DOI: 10.5287/ora-j2nr0wxob
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Subjects: Machine learning