Pushpak Pati

Since February 2022, I have been working as a Postdoctoral researcher at IBM Research Zurich. Prior to this, I obtained my Ph.D. under the supervision of Prof. Orcun Goksel in the Computer Vision Lab at ETH Zurich and Dr. Maria Gabrani at IBM Research Zurich.

Research highlights

• Graph Representation Learning: Graphs are flexible data structures to encode the phenotype and topological distribution of objects, e.g., cells in a tissue or atoms in a molecule, for contextually addressing downstream tasks, e.g., cell phenotyping or molecular property prediction. Motivated by their beauty, I exploited graphs for a variety of representation learning tasks:
  • comprehensive tissue microenvironment modeling via homogeneous and heterogeneous graphs
  • tackling giga-pixel image-level weakly-supervised classification and segmentation
  • interpretability and explainability of graph representations
  • multimodal integration using multiplexed graph representation
  I am the main contributor of HistoCartography, a library to facilitate graph analytics in digital pathology.

• Resource-Efficient Deep Learning: Typical deep learning models need massive amounts of labeled data for training. Inferring on them are also computation-intensive; hindering their deployment in limited infrastructure settings. In my research, I explored,
  • data-efficient methods: learn from sparse annotations
  • label-efficient methods: learn from inexpensive and readily available supervision
  • computation-efficient methods: deployable under computational constraints
  • generative modeling: synthesizing digital data for training augmentation

• Incremental Knowledge Refinement: Unlike humans, deep learning methods suffer from forgetting while continually learning new concepts, incorporating new data sources, and exploiting fine-grained information. To address these issues, I have been working on
  • continual unsupervised domain adaptation for classification and segmentation
  • differentiable zooming for multiple instance learning methods