Events Calendar

PhD Thesis Proposal - Zhuojin Li

Title: Performance Modeling and Optimization for Machine Learning Systems: from Cloud Training to Edge Inference
 
Date and Time: Fri, April 25, 2-4pm   
 
Location: EEB 403
 
Committee Members: Leana Golubchik (Chair), Murali Annavaram, Peter Beerel, Jyotirmoy V. Deshmukh, William G. J. Halfond
 
Abstract: Deep neural networks (DNNs) have achieved remarkable success in a wide range of tasks, from computer vision to natural language processing. However, as these networks substantially grow in scale, ensuring efficient performance across the entire lifecycle - from cloud-based training to edge-device inference - remains a crucial problem. Our work addresses this need by developing performance modeling and optimization techniques for both cloud-based distributed training and edge-based inference.  
 
First, we develop training throughput prediction models (coarse-grained and fine-grained) for distributed stochastic gradient descent (SGD), characterizing the impact of communication bottlenecks and node stragglers in synchronous/asynchronous and centralized/decentralized settings. Second, we propose an operation-wise framework that accurately predicts the inference latency of various neural architectures - such as CNNs and Vision Transformers (ViTs) - across diverse mobile platforms and ML frameworks. Finally, we propose a heterogeneous co-execution approach that combines low-overhead synchronization with ML-based workload partitioning on mobile CPUs and GPUs, substantially speeding up inference tasks. Together, these three contributions form a comprehensive methodology for end-to-end DNN performance evaluation and optimization, providing practical insights for large-scale training in the cloud and efficient deployment at the edge.

Location: Hughes Aircraft Electrical Engineering Center (EEB) - 403

Audiences: Everyone Is Invited

Contact: Zhuojin Li

This event is open to all eligible individuals. USC Viterbi operates all of its activities consistent with the University's Notice of Non-Discrimination. Eligibility is not determined based on race, sex, ethnicity, sexual orientation, or any other prohibited factor.

Add to Google Calendar