Jul 17 (Fri) @ 5:00pm: "Data-Efficient Machine Learning in Optimization and Circuit Performance Prediction," Yuxuan Yin, ECE PhD Defense

Location: Harold Frank Hall (HFH), Room 4110B
Zoom Link: https://ucsb.zoom.us/j/87486820686

Abstract

Many machine learning methods assume that additional labeled data can be collected whenever model performance is insufficient. However, in expensive black-box optimization and post-silicon circuit characterization, each label may require costly simulations, experiments, or hardware measurements. This dissertation develops data-efficient machine learning methods that maximize the value of each expensive observation by leveraging inexpensive auxiliary information. We first present a semi-supervised Bayesian optimization framework that improves optimization under limited evaluation budgets through optimized unlabeled sampling, teacher-student learning, and uncertainty-aware pseudo-labeling. We then introduce ADO-LLM, which combines large language models with Bayesian optimization to accelerate analog circuit design by incorporating design prior knowledge while grounding decisions with measured feedback.

The second part of the dissertation addresses reliable and data-efficient circuit performance prediction after fabrication. Focusing on minimum operating voltage Vmin, we develop conformal prediction methods that produce reliable prediction intervals using limited measurements and on-chip monitor data, and further propose methods that align inter- and intra-wafer variation to improve prediction accuracy while reducing characterization cost under distribution shift. Together, these contributions demonstrate that data efficiency can be achieved by integrating inexpensive auxiliary information with uncertainty estimation, calibration, and physical structure. Across both black-box optimization and semiconductor characterization, the proposed methods improve learning and decision-making while substantially reducing the need for expensive data collection.

Bio

Yuxuan Yin is a Ph.D. candidate in the Department of Electrical and Computer Engineering at the University of California, Santa Barbara. He received dual bachelor's degrees in Electronic Engineering and Mathematics from Tsinghua University in 2020. Since 2021, he has been pursuing his Ph.D. under the supervision of Professor Peng Li, with research focusing on data-efficient machine learning, Bayesian optimization, and electronic design automation (EDA). His work has been published at leading AI and EDA conferences, including ICML, DAC, ICCAD, and DATE. He previously interned with NXP Semiconductors in Austin, Texas, in 2023 and 2024, and with Google in 2025.

Hosted By: ECE Professor Peng Li

Submitted By: Yuxuan Yin <y_yin@ucsb.edu>