Fall 2025 GRASP SFI: Tianjiao Ding, University of Pennsylvania, “Learning Parsimonious Representations for Efficient Analysis and Synthesis”

This presenter is one of the winners of the 2025 GRASP vote for internal PhD or postdoc SFI Speakers!

This was a hybrid event with in-person attendance in Levine 307 and virtual attendance…

ABSTRACT

The automatic discovery of structures in data (analysis) and generation of data (synthesis) are two core problems in machine learning. Since data is high-dimensional and complex, a common paradigm is to learn a low-dimensional representation for data to facilitate both analysis and synthesis. However, existing methods are challenged by restrictive data assumptions and lack of semantic compositionality. We address these challenges by a unifying paradigm, which is to learn/leverage latent spaces supported on low-dimensional linear subspaces. Encoders and decoders then map between data and latent spaces. Such paradigm enables us to push multiple frontiers in data analysis and synthesis, including clustering images, aligning the semantics of text generation, and efficient image generation.