Self-supervised learning for Topological Neural Networks


Claudio Battiloro

Claudio Battiloro

Claudio Battiloro is a PhD candidate in ICT at Sapienza University of Rome, a Visiting Scholar at Harvard SPH, and a former Visiting Associate at the SEAS of University of Pennsylvania. He is a prospective co-appointed postdoctoral fellow at Harvard University and University of Pennsylvania. Claudio’s research interests include theory and methods for topological and algebraic signal processing, topological deep learning, and distributed optimization. He has several publications in top-tier journals and conferences. Claudio received different awards such as the IEEE SPS Italian Chapter Best M.Sc. Thesis Award (2020). In 2020, he graduated with distinction in the M.Sc. in Data Science with a (university-overall) Top 400 Students award at Sapienza University.


Self-supervised methods can be broadly categorized into contrastive approaches and predictive approaches. In the context of GNNs, self-supervision is particularly useful for learning graph encoders by utilizing information from the distribution of unlabeled graphs and minimizing a self-supervised loss. However, literature about self-supervised techniques for Topological Deep Learning is extremely sparse, and only a couple of works for simplicial complexes have been presented. For this reason, I propose to design novel self-predictive and contrastive objectives for training TDL architectures.

In particular, I will mentor the attendants with the aim of investigating generative objectives based on the predictive approach for architectures defined over regular cell complexes. We will explore cell complex-based reconstruction tasks that can properly generalize the well-studied node property prediction and attribute denoising tasks in the space of attributed graphs.

We will go through the actual SoA to design novel techniques able to combine my personal signal processing-grounded approach with cutting edge ML research on Topological and Geometric DL.