PhD defence: Neural Dimensionality Reduction for Data Visualization

PLEASE NOTE: The candidate gives a layman's talk, therefore the livestream will start fifteen minutes earlier.

Information is a crucial resource for humankind, as it allows us to achieve previously unimaginable goals. This work focuses on the study of information consisting of many related measurements per data point, which we call high-dimensional data. Being able to visualize and understand such high-dimensional data is key to the development and improvement of complex systems, such as machine learning models, but also to understanding patterns present in a given dataset when performing exploratory data analysis. One main approach to deal with this type of data is using Dimensionality Reduction (DR), a class of algorithms that maps the data-level problem of finding patterns in this complex data to a visual one. It does so by creating low-dimensional projections of the data, by preserving relevant attributes, and visualizing them in the form of a scatter plot. In this work, the synergies between neural networks, a popular, powerful, and versatile class of machine learning models, and DR techniques are explored, leading to novel techniques, analyses, and a deeper understanding of existing techniques, and how they are being evaluated.