Research

Three threads run through my work: making privacy guarantees survive contact with large models, making federated training cheap enough to be practical, and importing the structure-exploiting machinery of numerical analysis into places machine learning had not yet looked.

Multigrid–ViT Fusion for Seismic Foundation Models

AAAI'26 · CCF-A

Pioneered the first integration of multigrid numerical methods with Vision Transformers for seismic data, designing an adaptive two-grid training strategy that exploits the hierarchical structure in a seismogram.
Developed Hilbert curve encoding to preserve the spatial locality of 2D seismograms during tokenization, coupled with a Fourier-based spectral decomposition that separates low-frequency geological structure from high-frequency detail in embedding space.

96.4% reduction in MAE (0.1886 → 0.0068) against standard ViT baselines on foundation-model pretraining for seismogram reconstruction.

Communication-Efficient and Differentially Private Federated Fine-tuning for LLMs

AAAI'25, ECAI'25 · CCF-A/B

Developed CG-FedLLM (ECAI'25): low-rank gradient compression with learnable auto-encoders, achieving 98.44% gradient compression while preserving model performance on MMLU and C-Eval.
Pioneered DR-Encoder (AAAI'25): the first framework to jointly optimize gradient compression and differential privacy for LoRA fine-tuning with random Gaussian priors, reducing communication to 1.56% under strict privacy constraints.

Near-non-private accuracy on C-Eval (+1.2%) and MMLU (+0.8%) at ε = 0.25 — closing a utility gap where prior DP-LLM approaches (e.g. DP-FedLoRA) lose more than 50% at the same ε.

Differentially Private Knowledge Transfer for Recommendation

WWW'22 · CCF-A

Pioneered PriCDR-J and PriCDR-S, the first framework enabling cross-domain knowledge transfer under strict differential privacy guarantees (ε ≤ 0.5), overcoming the utility collapse inherent in naive DP applied to transfer learning.
Designed spectrally-aware perturbation mechanisms using randomized Hadamard transforms to preserve the low-rank structure of knowledge-transfer matrices; proved spectral distortion bounds via random matrix theory.

+8.43% accuracy over the private state of the art (PriCDR-SYM) on Amazon cross-domain benchmarks at ε = 0.5.

Why numerical analysis

Differential privacy and large-scale training are usually treated as separate engineering problems. They are both, at bottom, questions about how error propagates through an iterative process — which is the subject numerical analysis has studied for seventy years.

A preconditioner reshapes the spectrum of a problem so that an iterative solver converges faster; the same idea, applied to differentially private SGD, reshapes how injected noise interacts with the geometry of the loss. Multigrid exploits the fact that a signal decomposes across scales, and a seismogram or a galaxy field is exactly such a signal. Randomized sketching preserves the low-rank structure of a matrix under projection, which is precisely what a gradient-compression scheme needs to guarantee.

Working this way means the guarantees are provable rather than empirical, and the methods transfer between domains that otherwise share no vocabulary.