I develop mathematically grounded methods to make AI more data- and compute-efficient. My research connects three directions — data selection and mixing, optimization, and model design — with applications to language and multi-modal models.
Data selection & mixing
What should a model learn from, and when?
I study data for foundation models. I investigate how to combine training data across sources, semantic domains, and modalities such as text, images, and videos. I am also interested in selection methods that respond to what a model is learning and what it needs next.
Selected work: Chameleon · MaD-Mix · Multi-modal Spectral Clustering · MIA · OOD
Optimization
How can we make each update more effective?
I develop algorithms for faster model training and mathematical reformulations that make learning problems more efficient to solve. My work explores how to exploit problem geometry while accommodating requirements such as sparsity, robustness, and fairness.
Selected work: AcceleGrad# · Kernel PCA through dualization · Fair spectral clustering
Model design
Which computations and parameters does a model need?
I study architectures and parameterizations that reduce the cost of learning and inference. This includes efficient attention mechanisms, compact fine-tuning, and replacing expensive model computations with simpler approximations, while understanding the resulting trade-offs.
Selected work: LION · Quantum-PEFT · Neural Block Linearization · Deep Kernel PCA · Primal-Attention · IEEE Survey
My longer-term goal is to understand how data, optimization, and architectures can be designed together. I plan to investigate learning methods that expand the capabilities models can achieve with the available data and computational resources.