• Deep model architecture building experience

  • built the architecture of various generative models：BASNET, DCGAN, VQ-VAE, DANET, SPNET.

  • Used plug-and-play modules：Resblock, GhostBottleNeck, SE-layer, DarkBlock.

  • Used Attention mechanism：StripPooling, MixedPoolingModule, SelectiveKernel.

  • According to the input data, use convolutional layers of different dimensions (1D~3D) to learn information.

  • Used weight standardization to assign weights to improve model training effect.

  • Built a composite model of regression and classification.

• AI development environment management

  • Used docker or Anaconda to establish and maintain the development environment with GPU.

  • Set up the environment to use the LLM (LLAMA2, Taiwan-LLaMa, Codellama, Llama2-chinese-13b, etc.)

• Model Training and Tuning Tips

  • Adjusted the data batch size according to the hardware performance, and adjusted the normalization method in hidden layers.

  • Used Microsoft nni to adjust hyperparameters during model architecture and training.

  • Combined with Explainable AI methods in the training process.

  • Trained with Optuna and TPOT in machine learning projects.

• References Rewrite Schema 
  For deep learning projects, we referred to various literature and developed reusable modules that consistently improved model performance.
  • Examples include SelectiveKernel, GhostModule, MixedPoolingModule.

• Statistics Checking Skills

  • Regression model
    R-squared, RMSE, MAE, Residual Analysis, Correlation, POD, FAR, etc. 

  • Classification model
    ROC curve, AUC, Confusion Matrix, F1-score, recall.

Autotronic Enterprise Co., Ltd. (Aecl), May 2018 - Jun 2020

Data Analysis Engineer 

• Programming:

  • Designed anomaly detection programs for various instruments produced by the company.
  • Rewrote data encoding programs to ensure secure data transmission.
  • Visualized and generated necessary data for R&D and project requirements.
  • Conducted big data analysis using extensive instrument data with SQL and noSQL databases.

• Data Calibration - Machine Learning:

  • Integrated inspection through statistical tests and feature engineering.
  • Applied statistical models for quality inspection and utilized machine learning for data calibration.
  • Achieved over 90% accuracy in instrument data calibration using machine learning methods such as XGBoost, NGBoost, LightGBM.

• Web Scraping:

  • Developed web scraping programs using tools like Selenium and BeautifulSoup for machine learning data.
  • Utilized corresponding APIs for data retrieval and aggregation.

• Documentation:

  • Responsible for writing reports in proposals related to instrument comparisons and maintenance analysis.