Semi-Supervised Deep Learning with Class Imbalanced Data - SeDECI

Deep learning models are revolutionising automatic classification services in various sectors, including computer vision, medical diagnosis, and security. However, data annotation remains costly and time-consuming, prompting interest in semi-supervised as a solution.

Yet, guidelines for handling class imbalance, a common issue in real-world applications, are lacking within these frameworks.

This research aims to develop semi-supervised learning approaches that leverage large-scale data while effectively addressing class data imbalances in task-agnostic fashion.

The project focuses on a fundamental theoretical framework while experimenting with various techniques to find the best practices for specific use cases.

The time plan is set with three milestones: 1) theoretical study and identification of potential model architectures and techniques, 2) Development, optimization, and extensive evaluation of task-generic semi-supervised learning algorithms, and 3) Completion of the software modules, system integration, and further studies on model robustness and uncertainty quantification.This research primarily focuses on computer vision, such as object recognition and detection, fine-grained classification, anomaly detection, remote sensing and satellite imagery, and medical image analysis.

By developing task-generic semi-supervised learning algorithms capable of addressing data imbalance, this project seeks to contribute significantly to the field in real-world scenarios.