Wasing the always of wanting of knowing.
The eternal desire of a hungry soul is knowledge.
– Mistborn Series, Brandon Sanderson
My goal is to use theoretical and empirical insights to develop a principled understanding of machine learning (ML) models so that they are reliable when deployed in the real-world. More concretely, I am thinking about the following problems:
Graph embeddings distill high-dimensional feature and neighbourhood information onto low-dimensional spaces so that they can be used to perform ML tasks such as node classification and link prediction. I seek to formally explore the relationship between robustness and traditional notions of performance (e.g. accuracy), especially in the presence of random as well as structural perturbations to the input graph, and uncertainty in the form of noisy or unavailable data.
When we use graph embeddings for downstream tasks, it is imperative to ensure that they are systemically sound and safe to use. I am working on building models that are provably guaranteed to be robust and fair by design.
In the past, I worked on a related theme in the context of machine teaching. How can we construct useful training examples to simultaneously train stochastic gradient descent learners when we have incomplete or noisy information about them?
On the application side, I am excited about computational social science wherein I am interested in human behaviour on online social networks. My past research has explored how platforms facilitate trustworthiness amongst users through incentives and gamification. I have also studied the role of digital signalling on the actions, social capital, and social status of users online.
