Tamasha Malepathirana

PhD candidate

University of Melbourne

Tamasha Malepathirana is a 4th-year Ph.D. Candidate at the Faculty of Engineering and Information Technology at the University of Melbourne (UoM), supervised by Prof. Saman Halgamuge, Prof. Chennupati Jagadish, Prof Lezanne Ooi, and Dr. Vini Gautam. Her current research focuses on machine learning, with a keen affinity toward neural networks. Particularly, she is interested in the effective application of neural network algorithms in real-world applications and addressing the problems associated with them. Her research efforts have contributed significantly to the field, resulting in multiple journal and conference articles that are currently published or under review. In recognition of her accomplishments, Tamasha was awarded a GCI (Graeme Clark Institute) Women in STEM Student Award in 2022. Apart from her professional background, she was born and raised in Sri Lanka. In 2019, she moved to Melbourne, Australia, to pursue her higher education.

Neural Networks for High-Dimensional and Continually Generated Data

Neural Networks for High-Dimensional and Continually Generated Data Advancements in experimental technologies enable the generation of vast volumes of data in various fields, which, when coupled with the power of machine learning, lead to unprecedented breakthroughs. The analysis of this data is challenged by two phenomena.

Firstly, the "high-dimensionality" of datasets - a phenomenon characterized by a large number of variables. The low-dimensional visualizations of high-dimensional data provide a means to interpret intricate structures in them. However, the existing methods that generate visualizations struggle to capture discrete and continuous patterns simultaneously, thus her thesis investigates the ability of visualization methods in capturing both discrete clusters and continuous progressions within high-dimensional data.

Secondly, continual generation of experimental data, i.e., data generated over time instead of being available all at once, requires machine learning models that can be regularly updated to remain relevant to newly generated data. This is crucial in longitudinal studies where the objective is to capture the temporal variations in the measured variables accurately. Therefore, her thesis proposes a continually trained visualization pipeline that can effectively integrate newly generated high-dimensional data into existing low-dimensional visualizations.

Finally, she presents a novel technique to address a significant research gap in the continual training of neural networks. Specifically, the reliance on previously encountered data, which can be impractical in applications where storing old data is not viable. Altogether, her thesis contributes to the effective application of machine learning in real-world applications, particularly in the areas of visualization and continual learning.