Research

My main research goal is to facilitate deep learning on structured data, such as sets, point clouds, and graphs, by developing new theoretical frameworks and practical algorithms.

These efforts are driven by the growing number of applications involving such data. For example, 3D shape analysis, social network analysis, and molecular property prediction.

Applying deep learning to these data types poses distinct challenges, since they have inherent symmetries: sets are invariant to the order of their elements, graphs to relabeling of their nodes, and point clouds to rotations and translations. Consequently, naively feeding such data to neural models typically results in overfitting irrelevant aspects of the input representation. For instance, in the case of sets, the model may overfit the order in which the set elements are given.

While it is relatively easy to construct architectures that are invariant to these symmetries (e.g., sum-pooling in the case of sets), a key challenge is that imposing invariance often comes at the cost of expressiveness. To unlock the full potential of deep learning for these data types, it is therefore necessary to develop new theory and algorithms.

Our group’s efforts have already produced several results addressing these challenges, published in top-tier venues. For example:

Recent work by independent groups has corroborated that the FSW embedding’s metric properties can translate to improved performance on practical learning tasks (see, e.g., Shivottam and Mishra (2026) and Chen et al. (ICLR 2026)).

We are currently working on extending our approach to 3D point clouds under rotation invariance, with applications to molecular data.