Shengming Zhang

Implemented the HMCN model per Tencent's guidelines on the PaddlePaddle Chinese spam SMS dataset. Developed a Python script to label 2,222 messages with ERNIE in 17 subcategories across 6 majors, saving and extracting from CSV to create a JSON dataset. Achieved 95% testing accuracy after 50 rounds of TextCNN training.

Developed and optimized DIP and 3DUnet image denoising techniques on Alibaba Cloud, achieving optimal stopping condition by fine-tuning the learning rate to LR=6e-3 for convergence at 600 epochs. Defined ROI by isolating gray matter in simulated datasets, using it as a standard to calculate CNR and determine the optimal denoising endpoint

Led the development of conversational emotion recognition for human-computer dialogue systems using the CPED dataset, achieving an accuracy rate of 24.14% which surpasses the human average of 23.04%. Solve AI alignment issues, designed training data excluding personality traits and dialogue behaviors, and created 4 test plans with 2 test cases.