Priyadarshini Kumari

I build machine learning systems that have to work when the data, supervision, or world itself is a little less cooperative than the benchmark would suggest :)

I currently work on the Siri AI team at Apple, where I focus on making Siri better at answering personal questions - connecting models, retrieval, reasoning, and personal context while keeping the experience useful, reliable, and appropriately private. Read more about the work in Apple’s announcement.

Previously, I was part of the Apple Health team, where I worked on machine learning models and agentic systems for health applications. Before Apple, I was at Sony AI, where I worked on data-efficient and multimodal perception learning. My research included learning with limited supervision, graph neural networks, multimodal perception, and models that connect very different kinds of information—including text, vision, and olfactory signals. The applications ranged from biomedical research to olfaction and gastronomy.

I received my Ph.D. from IIT Bombay, advised by Prof. Subhasis Chaudhuri and Prof. Siddhartha Chaudhuri. My thesis, Label-Efficient Distance Metric Learning, focused on learning useful representations and similarity functions when labeled data is scarce. Before my Ph.D., I completed my master’s at IIT Bombay, where I worked on multimodal rendering — combining haptic, visual, and auditory feedback to make 3D models of heritage sites more accessible to people with visual impairments.

Across these projects, I keep coming back to a fairly simple question: How do we make machines learn something useful without requiring the world to first become a perfectly labeled dataset? I am still working on that. The datasets have, so far, declined to cooperate.