Research Experience
University of Southern California
Research Assistant, Aug. 2017 - Present
Advisor: Prof. Bhaskar Krishnamachari
Task scheduler for dispersed computing systems: Proposed a novel scheduling algorithm for high through- put task execution in the dispersed computing environment, and integrated the scheduler into the state- of-the-art Jupiter orchestrator. Addressed performance bottlenecks via task duplication and splitting.
Fast and accurate streaming CNN inference via communication compression on the edge: Proposed a streaming inference framework to simultaneously improve throughput and accuracy based on novel communication compression techniques. Designed an encoder-decoder architecture to significantly com- press the data traffic among collaborating edge devices. Implemented an adaptive compression scheme based on quantitative tradeoff analysis among compression ratio, throughput and accuracy.
Intelligent beam selection for 5G vehicular networks: Proposed a reinforcement learning model to en- able online selection of millimeter-wave beams by the base stations. Designed a Deep Q-Learning Network architecture to improve data rate adaptively according to road and traffic conditions.
University of Southern California
Graduate Level Course, Mar. 2019 - June 2019
Instructor: Prof. Joseph J. Lim
ReGAiN: Recommendation with Graph Adversarial Nets: Captured user-user relations in social rec- ommendation systems with Graph Convolutional Network (GCN). Designed and implemented a novel content recommendation model for social networks by combining the advantage of Graph Convolutional Network (GCN) with Generative Adversarial Network (GAN). Used adversarial training to promote recommendation quality over traditional methods such as Collaborative Filtering.
University of Southern California
Research Assistant, Aug. 2016 - Aug. 2017
Advisor: Prof. Viktor Prasanna
Accelerating all-pair-shortest-path computation on FPGAs: Improved throughput and energy efficiency of the FPGA accelerator executing the Floyd-Warshall algorithm. Proposed and implemented a systolic array-based pipeline architecture. Developed a performance model for the optimal mapping parameters.
Research Activities
University of Southern California
Paper Reviewer
Heterogeneous High-performance Reconfigurable Computing (H2RC) Workshop, 2016
IEEE Field-Programmable Custom Computing Machines (FCCM) Conference, 2017
IEEE International Conference on Big Data (BigData), 2019
IJCAI, 2020
Publications
Diyi Hu, Bhaskar Krishnamachari, Fast and accurate streaming CNN inference via communication compression on the edge, ACM/IEEE conference on internet of things design and implementation (IoTDI), 2020 (pdf, to appear)
Diyi Hu, Bhaskar Krishnamachari, Throughput optimized scheduler for dispersed computing systems, 2019 7th IEEE International Conference on Mobile Cloud Computing, Services, and Engineering (MobileCloud), 2019 (pdf)