Exploring Probabilistic Models for Semi-supervised Learning

TL;DR

The thesis tackles the challenge of semi-supervised learning (SSL) under uncertainty by developing probabilistic SSL methods that provide reliable uncertainty estimates while maintaining competitive accuracy. It introduces a fully Bayesian Generative Bayesian Deep Learning (GBDL) framework for semi-supervised volumetric medical image segmentation, and two Neural Process–based SSL mechanisms: NP-Match for large-scale image classification and NP-SemiSeg for semi-supervised semantic segmentation. A key innovation is the uncertainty-guided skew-geometric Jensen-Shannon divergence $JS^{G_{oldsymbol{\alpha_u}}}$, which improves robustness to noisy pseudo-labels and enhances uncertainty calibration. Together, these methods yield faster and more reliable uncertainty quantification, enabling safer deployment in safety-critical domains such as healthcare and autonomous systems, and demonstrate strong empirical performance on standard SSL benchmarks and medical imaging datasets. The work lays a foundation for broader adoption of probabilistic approaches in SSL and points to future avenues in open-set SSL, translation-equivariant neural processes, and scalable Bayesian SSL architectures.

Abstract

This thesis studies advanced probabilistic models, including both their theoretical foundations and practical applications, for different semi-supervised learning (SSL) tasks. The proposed probabilistic methods are able to improve the safety of AI systems in real applications by providing reliable uncertainty estimates quickly, and at the same time, achieve competitive performance compared to their deterministic counterparts. The experimental results indicate that the methods proposed in the thesis have great value in safety-critical areas, such as the autonomous driving or medical imaging analysis domain, and pave the way for the future discovery of highly effective and efficient probabilistic approaches in the SSL sector.

Figures (19)

• Figure 1: Overview of the SSL pipeline. Unlike active learning, SSL does not require user involvement in choosing data. The selected data are used to enlarge the original labeled dataset, and whether pseudo-labels are used depends on the training strategy employed (e.g., pseudo-labeling strategy or consistency strategy).
• Figure 2: Overview of the original NPs.
• Figure 3: GBDL for semi-supervised volumetric medical image segmentation, including a latent representation learning (LRL) architecture (in the green dotted box) and a regular 3D-UNet with MC dropout (in the red dotted box). Only the regular 3D-UNet with MC dropout is used during testing. For simplicity, the shortcut connections between the paired 3D-UNet encoder and decoder are omitted.
• Figure 4: The influence of the number of feedforward passes on: (a) time costs (ms) for processing a $128\times128\times32$ volume input on a single GeForce GTX 1080 Ti; (b) Dice score; and (c) PAvPU metric. The horizontal axis refers to the number of feedforward passes, namely, $T$.
• Figure 5: The visualization of some predicted slices from the KiTS19 dataset for different Bayesian deep learning based methods yu2019uncertaintywang2020doublewang2021tripled.