Research Projects 

My third research thrust extends efficient on-device AI from inference and model selection toward resource-efficient model adaptation. I developed energy- and memory-efficient PEFT methods for personalized SLMs, systematically evaluating adaptation strategies across Transformer- and Mamba-based models under constrained computing resources [arXiv 2026a]. I then extended this direction to vision models and VLMs by combining PEFT with adaptive activation checkpointing, enabling fine-tuning under explicit memory constraints on resource-constrained GPUs [arXiv 2026c]. Complementing these adaptation techniques, my GPU-DVFS work demonstrated that the energy cost of on-device fine-tuning can also be reduced through adaptive hardware operation [arXiv 2026b]. Across these studies, I used NSE, NSM, and NS# to characterize the tradeoffs among predictive performance, energy consumption, and peak memory during model adaptation. Together, this work establishes a foundation for memory- and energy-efficient model adaptation on resource-constrained platforms and motivates my future research on joint model--system optimization.

Building on the ML-enhanced CPU–GPU DVFS governor introduced in my PhD work [ML-Gov: ESTIMedia 2017], I extended this integration of embedded systems and machine learning toward systematic model selection and interpretable runtime decision-making [ACM TECS 2021]. In this journal extension, I investigated interpretable predictive models and workload descriptors, including simulatability operation counts (SOC), to capture workload-dependent power–performance behavior and dynamically scale CPU and GPU frequencies for improved energy efficiency. I subsequently explored offline reinforcement-learning-oriented objective-function design for dynamic frequency scaling of CNN workloads [BigComp 2023a] and integrated CNN workload characterization with CPU–GPU DVFS governors on embedded platforms [IEEE ESL 2023]. More recently, I extended adaptive runtime resource management to on-device AI training by developing GPU-DVFS techniques that reduce the energy cost of fine-tuning small language models (SLMs) on resource-constrained embedded platforms [arXiv 2026b]. Collectively, this thrust extends my earlier CPU–GPU power-management research toward adaptive resource management for modern AI workloads, with a focus on improving energy efficiency while maintaining model performance.

As a graduate student at UCI, I focused on energy-efficient runtime management for emerging heterogeneous mobile GPU systems, particularly on coordinating CPU, GPU, and memory resources under dynamic workloads. 1) First, I characterized mobile graphics workloads from a quality-aware perspective to understand the relationships among performance, visual quality, and energy consumption [ESTIMedia 2014]. 2) Second, I developed two complementary runtime power-management approaches: cooperative CPU–GPU DVFS, including memory-aware control [ESTIMedia 2015; DAEM 2018], and cooperative and hierarchical frequency capping for energy-efficient mobile gaming [Co-Cap: SAC 2016; HiCAP: ISLPED 2016; ACM TECS 2018]. 3) Finally, I extended these adaptive runtime techniques with a machine-learning (ML) enhanced CPU–GPU DVFS governor using model trees that learned workload-dependent power–performance behavior to improve runtime decision-making [ML-Gov: ESTIMedia 2017]. Collectively, this work established my foundation in adaptive resource-aware computing and led naturally to my later research on efficient on-device AI.