Sourya Dey
Sourya Dey photo

Sourya Dey


AI researcher,
sports enthusiast.

sourya.dey@gmail.com
sourya@galois.com

Linkedin
Google Scholar

Professional Biography and Research Interests

Sourya Dey is currently working as a Research Engineer at Galois in the Washington DC metro area. He completed his PhD specializing in machine learning at the University of Southern California (USC) in 2020, where he was a recipient of the Viterbi Graduate School PhD Fellowship. Prior to that, he obtained his Bachelor of Technology (Honours) degree in Instrumentation Engineering from the Indian Institute of Technology (IIT) Kharagpur in 2014, along with the Best B. Tech Project award in his department.

My research interests are primarily machine learning and artificial intelligence (ML/AI), with recent focus on interpreting large language models (LLMs). I am interested in the applications of ML in various fields such as modeling systems, formal verification, and privacy and security. This is reflected by my work at Galois, which involves interdisciplinary projects at the intersection of machine learning, data science and software engineering. As one example, we have designed a Python package DLKoopman to use deep learning to model and predict the behavior of dynamical systems such as pressure on an aircraft.

My PhD research can be summarized by the title of my thesis — Exploring Complexity Reduction in Deep Learning. This included developing the pre-defined sparsity technique to simplify deep learning architectures with minimal performance degradation, and the open-source automated machine learning (AutoML) framework Deep-n-Cheap to search for deep learning models and explore tradeoffs between performance and complexity. I was involved in the inception and subsequent growth of a new machine learning group at USC, and also received a Best Paper award for developing a family of open-source synthetic machine learning datasets.

Recent News Sep '25: Our paper Token embeddings violate the manifold hypothesis got accepted for publication at NeurIPS. arXiv version here, camera-ready conference version coming soon!
Jul '25: Our paper A Simulated Reconstruction and Reidentification Attack on the 2010 U.S. Census got accepted for publication at Harvard Data Science Review (HDSR).


DLKoopman logo

Deep-n-Cheap logo

A bit about my personal side

I grew up in the beautiful and very unique city of Kolkata -- the City of Joy. My undergrad days were in nearby Kharagpur, before moving to the entertainment capital of the world -- Los Angeles. I currently live in Arlington in the Washington DC metro area. The natural beauty here is amazing, lots of opportunities to bike and hike!

I like reading, particularly detective thrillers, science fiction and fantasy, and sports-related books. Here's my Goodreads profile, I'd love to connect with other bibliophiles! I am also a big fan of watching football. My ideal weekend is waking up in time to watch Chelsea play and (hopefully) win. Some other activities which I enjoy are swimming, coding, and experiencing new places and cuisines.

Guinea pig dinner in Peru
Guinea pig dinner in Peru