Pingyuan Liang Email & Phone Number

United States

United States

San Jose, California, United States

• Vehicle Intent Prediction
• Model releases: v8 model, responsible for the offboard part, uses convolution to implement time series modeling, performs rotation alignment, and greatly improves the accuracy on the benchmark, with the classification.
• Screening and fixing old bugs: Check and fix multiple bugs that have a significant impact on model performance, including feature confusion caused by improper cache cleaning; outliers caused by unclear category.
• Others: Reformatting of the code base, removing a lot of hard code; visualization tools to facilitate viewing input and output statistics;

California, United States

• Vectorized input for a large language model (LLM) to establish a novel architecture, significantly enhancing its performance.
• Implemented advanced prompt engineering, RCI techniques, and chain of thoughts strategies, boosting overall performance by a significant 1.5%.
• Developed a strategy for compliance loss-based command. Resulting in better command compliance.

Pittsburgh, Pennsylvania, United States

Construct anchor-based trajectories, gain 1% accuracy; refine for better representation, gain another 1% accuracyConstruct raster Input, decoupled objects, like boundaries and traffic lights, gain 2% accuracyBuild lane raster by gradient and angle, to include direction information, gain 1% accuracyBuild a occ-flow model, achieve 1% accuracyWin the Second.

Xiaomi Science Park, Haidian District, Beijing

Translated a mass production model from MXNet to PyTorch, achieving an accuracy alignment within 0.01%Developed an auto pipeline for dataset preparation, train job, model saving, model inference and model evaluation, implementing multi-processingDeveloped in a framework for integrating datasets, speeding dataloader/trainer pipeline, camera geomotry model.

Fangshan District, Beijing, China

Design a taillight detection model based on center point detection and a classification model based on multi-layer feature fusionUse an object-tracking model with Kalman filter and Hungary algorithm to track vehiclesAdopt time-series model GRU to make decision from tracking and classification resultAchieve a great accuracy of taillight classification of.

Shanghai, China

Design object detection network based on RetinaNetImplement multi-feature fusion from different layers based on the DarkNet/EfficientNet backboneDesign several headers on different feature layers, plan the data collection process and adjust the training processAchieve an AP of more than 90% on traffic light detection

Master'S Degree, Electrical And Computer Engineering

Bachelor'S Degree, Mathematics And Applied Mathematics

Bachelor'S Degree, Psychology