Lei Zhao （赵雷）

• Identified an optimization opportunity where interconnect bandwidth (NCCL) remains idle during forward pass, allowing for latency hiding of dynamic feature transmission.
• Designed a "Planner-Syncer" collaborative system to exploit this idle time. The Planner pre-calculates data placement (GPU-retention vs. CPU-writeback), enabling the Syncer to asynchronously multicast next-batch features in parallel with the current forward pass.
• Devised a hybrid synchronization strategy to mitigate Python GIL contention: adaptively utilizes P2P transport for high-dimensional features to maximize throughput, while falling back to shared memory for low-dimensional features to minimize overhead.

Key technologies: PyTorch Distributed, Caching, DGL, NCCL, GLOO, shared memory, multi-process, multi-thread, zero-copy

Lei Zhao, Longsheng Wei, Zhi Ma, Zhiheng Liu. Multimedia Tools and Applications, 84(19):20995–21016, 2025.

• Design the network structure based on DCN and customized UperNet
• Reduce the UperNet layers and simplify the PPM head as pavement crack detection relies on local textual details rather than global semantic information.
• Adopt deformable convolution as pavement cracks are usually strips rather than square
• Fuse features from different heads and the weight of each channel is trainable

Zhi Ma, Lei Zhao, Longsheng Wei. Journal of Advanced Computational Intelligence and Intelligent Informatics, 27(4):561–566, 2023.

• Design the network structure based on Cross-Attention and Resnet-101
• Add edge sensing module for better perception of edge

Longsheng Wei, Lei Zhao, Jian Peng. Applied Sciences, 11(20):9776, 2021.

• Design the network structure based on SIFT and Saliency features

• Purchased hardware (motherboard, PSU, RAM, GPU, PCIe Splitter, SSD, HDD, etc.)
• Installed Xubuntu and configured CUDA, Docker, rule-based proxy, NAT traversal, X11 forwarding, etc.
• Setup dataset server and shared directory with NFS
• Composed Dockerfile facilitating the quick deployment of frequently-used environments
• Carried out benchmarks on various combinations of hardware (Intel processor E5v3, E5v2, i7 family, PCIe x1 x4 x8 x16, PCIe 1.0 and 3.0, Nvidia 1080ti, 2080ti, 3090, P102-100, P106-100, etc.)
• Performance tuning with htop, nvtop, iotop, top, bcache and swap

• Developed a module to translate MATLAB data to Python pickle dump
• Designed a CNN-based network to fuse EEG and EOG features for real-time driver fatigue detection
• Prepared slides and poster for demonstration

• Developed the image preprocessing module to combine the original image and generate images of target areas according to carriage characteristics
• Adopted YOLOv5 and FastRCNN in the malfunction detection process
• Developed a module to fuse YOLOv5 and FastRCN recognition results and provide Flask API
• Composed Docker Compose file facilitating the instant deployment of all modules (frontend, java-backend, python-backend, database)

• Developed websites, chrome extensions, and WeChat applets

• System Integration
• Network Optimization
• Signals Analysis and Feature Verification against 3GPP Specs

• Participated in the revision of university BBS
• Developed a website named 59score by reverse-engineering the undergraduate registrar system, enabling students to check their grades.