Learning from limited labeled data
File(s)
Author(s)
Lazarou, Michalis
Type
Thesis
Abstract
A fundamental bottleneck of deep learning models that inhibits their use in many real world applications is their reliance on large labeled datasets. This thesis addresses this limitation by proposing several ideas that can improve the performance of deep learning models when the available labeled data is limited. The thesis proposes a novel algorithm for semi-supervised and transductive inference when the labeled data is limited achieving state of the art performance in the process. The proposed algorithm leverages the manifold structure of the labeled and unlabeled data to assign accurate labels to the unlabeled data before a label cleaning module is used to select the most confident pseudo-labeled data to be used as labeled data and augment the labeled dataset. Additionally, in real world situations the unlabeled data may contain noisy images. Using the unnormalized manifold similarity obtained from label propagation along with the label cleaning module, the aforementioned problem can be addressed effectively and filter the noisy data. Moreover, the thesis identifies a limitation of the traditional label propagation algorithm and proposes a novel variant that assigns more accurate labels to the unlabeled data. The idea is to iteratively optimize the positions of the labeled data in the manifold by minimizing a differentiable loss function. The thesis also investigates the performance of deep learning models when the available unlabeled data is class-imbalanced. A novel algorithm is proposed that achieves state of the art performance by combining effectively the two main lines of research in transductive learning for limited labeled data, which are: class centroid refinement and data manifold exploitation. Lastly, in certain applications obtaining data is extremely difficult. A novel generative model is proposed to generate synthetic data and train a classifier on both real and synthetic data.
Version
Open Access
Date Issued
2023-11
Date Awarded
2024-02
Copyright Statement
Creative Commons Attribution NonCommercial NoDerivatives Licence
Advisor
Stathaki, Tania
Publisher Department
Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)