Riyasat Ohib
I’m a Research Scientist at Google DeepMind in New York City. I got my PhD from Georgia Tech, where I was advised by Vince Calhoun and Sergey Plis.
Currently, I’m interested in building and understanding language and audiovisual models. How can we learn representations that bring language and audiovisual information into a coherent model of the world? How can that lead to more reliable reasoning and generation, and safer, more responsible AI systems? More broadly, I’m interested in how human learning and perception can inform the design of models and agents. During graduate school, my research focused on sparsity in deep and reinforcement learning, including work during research internships at Meta FAIR, Google DeepMind, Cohere, and Dolby Laboratories.
When I’m not doing research, I’m usually playing soccer, traveling, hiking, or reading about science, history, and how the world works. I keep a list of what I’ve been reading.
Selected work
-
Concept Spaces in the Residual Stream of Diffusion Transformers
Training-free steering of diffusion transformers by intervening on concept directions in the residual stream, often matching or beating prompting with smoother control.
-
SSFL: Discovering Sparse Unified Subnetworks at Initialization for Efficient Federated Learning
SSFL finds sparse subnetworks before federated training starts, reducing communication while preserving accuracy across non-IID clients.
-
Exploring Sparse Adapters for Scalable Merging of Parameter Efficient Experts
Sparse adapters provide simple, parameter-efficient task experts that merge more effectively than LoRA or full fine-tuning across as many as 20 NLP tasks.
-
Efficient Reinforcement Learning by Discovering Neural Pathways
Neural pathways identify compact subnetworks for reinforcement learning, retaining strong performance with less than 5% of a larger network's parameters and supporting multiple tasks within one model.
News
-
Presented SSFL at the Flower Monthly Research Seminar, hosted by Flower Labs. Watch the recording.
-
Joined Google DeepMind as a Research Scientist in New York, working on understanding representations and how to control frontier multimodal models.
-
Completed my PhD at Georgia Tech with Vince Calhoun and Sergey Plis. Dissertation: The Hidden Structure of Deep Models: Discovering and Exploiting Sparse Subnetworks.
-
Learning to Query History accepted at the ICLR TSALM workshop, on retrieval-augmented classification under nonstationarity.