System-Efficient Federated Learning (FL) on Heterogeneous Systems

• Knowledge distillation for exploiting heterogeneous edge devices in FL

• Partial model aggregation for communication-efficient FL

• System-aware partial model training for FL on heterogeneous systems

Distributed Optimization Algorithms for Large-Scale Deep Learning

• Computation-efficient Sharpness-Aware Minimization (SAM)

• Gradient recycling for communication-efficient deep learning

• CPU-aided asynchronous optimization for large-scale deep learning

Applied Machine Learning / Deep Learning for Electronic Materials Design and Analysis

• Deep learning-based Power Spectral Density (PSD) analysis for estimating the oxygen vacancy distribution in heterostructures.

Large-scale communication-efficient distributed learning (@ USC)

• Partial model training: theoretical analysis on the impact of partial model training

• Layer-wise adaptive model aggregation strategy for scalable Federated Learning

• Hessian-aware learning rate adjustment method for large-batch neural network training

Scalable parallelization strategy for deep learning on HPC platforms (@ Northwestern)

• Communication-efficient neural network training by exploiting the overlap of computation and communication

• Gradient averaging algorithms with a lower communication complexity

• Adaptive model update frequency control for scalable deep learning

I/O strategies for large-scale deep learning on HPC platforms (@ Northwestern)

• Asynchronous I/O strategy for data feeding in deep learning

• Parallel HDF5 data analysis