I am particularly interested in developing a scientific understanding of deep learning methods, with the goal of making them more accessible and efficient. In particular, I study how different inductive biases—such as data structure, scheduling, training algorithms, and model architecture—shape learning dynamics and ultimately affect performance.