Author ORCID Identifier
Date of Award
Doctor of Philosophy (PhD)
The success of Deep Learning in various tasks is highly dependent on the large amount of domain-specific annotated data, which are expensive to acquire and may contain varying degrees of noise. In this doctoral journey, our research goal is first to identify and then tackle the issues relating to data that causes significant performance degradation to real-world applications of Deep Learning algorithms.
Human Activity Recognition from RGB data is challenging due to the lack of relative motion parameters. To address this issue, we propose a novel framework that introduces the skeleton information from RGB data for activity recognition. With experimentation, we demonstrate that our RGB-only solution surpasses the state-of-the-art, all exploit RGB-D video streams, by a notable margin.
The predictive uncertainty of Deep Neural Networks (DNNs) makes them unreliable for real-world deployment. Moreover, available labeled data may contain noise. We aim to address these two issues holistically by proposing a unified density-driven framework, which can effectively denoise training data as well as avoid predicting uncertain test data points. Our plug-and-play framework is easy to deploy on real-world applications while achieving superior performance over state-of-the-art techniques. To assess effectiveness of our proposed framework in a real-world scenario, we experimented with x-ray images from COVID-19 patients.
Supervised learning of DNNs inherits the limitation of a very narrow field of view in terms of known data distributions. Moreover, annotating data is costly. Hence, we explore self-supervised Siamese networks to avoid these constraints. Through extensive experimentation, we demonstrate that self supervised method perform surprisingly comparative to its supervised counterpart in a real world use-case. We also delve deeper with activation mapping and feature distribution visualization to understand the causality of this method.
Through our research, we achieve a better understanding of issues relating to data-driven learning while solving some of the core problems of this paradigm and expose some novel and intriguing research questions to the community.
Sarker, Krishanu, "Data Efficient Learning: Towards Reducing Risk and Uncertainty of Data Driven Learning Paradigm." Dissertation, Georgia State University, 2021.
File Upload Confirmation