About Me
I'm a 3rd year PhD student in Computer Science at the University of Wisconsin-Madison, working with Professor Vikas Singh. I graduated from Carleton College with degrees in Mathematics and Computer Science. I also studied abroad at DIS Copenhagen and Budapest Semesters in Mathematics.
My research is in machine learning, and I aim to develop mathematical foundations for building more efficient and reliable AI systems. I'm interested in understanding the compositional structure of deep models and ensuring they behave predictably in real-world settings.
I'm from Japan, and my name is written 山田大介. Outside of research, I recently started learning piano and am enjoying the learning process :)
Research Projects
Composing Linear Layers from Irreducibles arXiv NeurIPS
We introduce a Clifford algebra-based decomposition of linear layers into geometric primitives with a differentiable algorithm. Applied to LLM attention projections, the method reduces parameter counts from O(d²) to O(log²d) while maintaining baseline accuracy and perplexity. We achieved this exponential reduction in raw parameter counts through decomposing the behaviors of dense linear layers into a substantially smaller set of bivectors that collectively define rotations.
Adaptive Scoring and Thresholding for Robust OOD Detection arXiv
We develop a mathematically grounded OOD framework for adaptively updating scoring functions and thresholds with human feedback while maintaining false positive control. The framework provides mechanisms for adapting to OOD inputs under dynamic conditions.
Is Conformal Factuality Robust to Distractors?
We investigate the robustness of conformal prediction-based guarantees for LLM factuality under realistic distribution shifts. We demonstrate that these guarantees can break down in the presence of distractors, highlighting reliability challenges in AI systems.
Image-Adaptive Generative Adversarial Networks PDF
We replicate Image-Adaptive GANs for image reconstruction tasks and build generative models capable of accurately estimating compressed or degraded images with high confidence. We analyze the model performances against images with different facial features to identify biases. This project was done as part of my undergraduate thesis at Carleton.
Selecting Balls From Urns With Partial Replacement Rules Paper
We investigate the "preferential" distribution X(0,1) in combinatorial probability. We explore open questions regarding the distribution including density, expectation, and variance properties.
Contact
Feel free to reach out - I'm always happy to chat about research or potential collaborations!