I am a Ph.D. student in ECE at UC Santa Barbara, advised by Prof. Zheng Zhang. I work on memory- and energy-efficient training.
My core direction is back-propagation-free (BP-free) training: gradients are estimated from forward passes alone (zeroth-order optimization), which makes on-device training a plug-in over an inference accelerator: a training engine needs little more than the inference engine already on the chip. I apply it to:
I also work on efficient LLM fine-tuning and optimizers (FuRA, Muon²), and have interned at Amazon Lab126 and Hewlett Packard Labs.
TL;DR LoRA-level cost, Full-FT-level performance (+1.37 over Full FT on LLaMA-3-8B). Each layer is factored as W = L·S·R with L fixed to the pretrained SVD basis; training only S and R gives full-rank, spectrally preconditioned updates.
TL;DR Add a control unit and repeated forward passes to a quantized inference engine and it becomes a training engine. Full-model training fits in 1 MB SRAM (sparse: 256 KB) and matches BP accuracy.
ACM TODAES 2025 →
TL;DR A sparse-grid Stein estimator removes BP from the PINN loss, and tensor-train ZO training keeps the search space small. Together they enable real-size PINN training on photonic tensor cores.
ACM TODAES 2026 →★ first-author work. * equal contribution.