Mitigating data scarcity in real-world problems

Abstract

Data scarcity is one of the prominent challenges in training deep neural networks. This dissertation investigates the critical challenge of data scarcity for real-world applications, where data can be scarce, unbalanced, and distributed irregularly over time or across domains. It presents novel methodologies to mitigate the issues of time-level and class-level data scarcity which impede the effectiveness of deep learning models. To address the temporal aspect of data scarcity, we leverage the potential of neural ordinary differential equations for generating frames in video data, enabling smoother and continuous video generation. Moreover, we propose innovative ensemble and adapter techniques to combat class imbalance problem, ensuring more robust and equitable performance across classes with varying instance frequencies. These approaches are geared towards improving model performance in various real-world scenarios. The dissertation promises a comprehensive exploration of these methods, aiming to significantly enhance the capabilities of deep learning in data-constrained environments.