| dc.description.abstract | During the past couple of decades, machine learning and deep learning methods have
achieved remarkable results in many real-world applications. However, it is difficult
to develop and train these artificial intelligence algorithms without a labeled dataset.
Under this circumstance, it is desirable to leverage a large number of unlabeled data
into the training process with fewer or even without labels. To this end, a non-supervised learning strategy (e.g., unsupervised, semi-supervised, weakly-supervised,
or self-supervised) has recently been studied in different domains.
In chapter 3, a novel semi-supervised framework is proposed to solve a clustering problem fundamentally by involving only few numbers of labeled data. In this
proposed framework, a non-iterative autoencoder is proposed for learning a representation of each data in an unsupervised way. The experimental results theoretically
demonstrate the effectiveness of this proposed framework, where the obtained clustering accuracy for thirteen tabular and image datasets are impressive. It has also
shown that the proposed autoencoder is able to capture important features of each
data.
In chapter 4, the above framework is extended to a weakly-supervised semantic segmentation task for demonstrating its practical ability. Before applying the
modified proposed framework to this task, computer vision methods are presented
as preliminary work to generate the initial labeled data and clustering space. We
achieve the current state-of-the-art performance on PASCAL VOC 2012 dataset.
This thesis shows that the proposed framework is capable not only for the traditional machine learning problem but also for the widely used real-world applications. | en_US |
