cv
General Information
| Full Name | Kelvin Ng |
| Date of Birth | 7th May 1997 |
| Languages | English, Cantonese, Mandarin |
Academic Interests
-
Representation Learning
- Interpretable representation learning methods for building so-called world model/system identification
- Continual learning of representation for building world model
- Computer vision based representation learning
- How spatial semantics and temporal structure of data emerge
-
Reinforcement Learning/Control Theory
- Optimal control with learned world model
- How optimal control task may impact representation learning
Education
-
2019 BSc in Mechanical Engineering
The University of Hong Kong - 1st Class Honours.
- Microsoft Innotech Law Hackathon 2018, First Runner-up
- Fuzzy system and neural network (A-)
- 1st Class Honours.
-
2023 Data Science Summer Camp - Theory, Algorithms and Applications
Institute of Data Science, The University of Hong Kong - Machine learning performance evaluation, Bayesian classifier, linear programming, nonlinear programming, network theory
-
2022 Reinforcement Learning Specialization - Online Course
University of Alberta - Understand the space of RL algorithms (TD learning, Monte Carlo, Sarsa, Q-Learning, Policy Gradient, Dyna and more)
- Implement RL algorithms (TD, Expected Sarsa, Q-Learning, Dyna, Actor-Critic) on toy datasets, in both discrete and continuous state environments
-
2022 Natural Language Processing with Sequence Models - Online Course
Deeplearnging.ai - Learn about neural networks for sequential data modeling (RNN, LSTM, GRUs)
- Implement sequential model on different tasks (Named Entity Recognition, Next-Word Generator, Sentiment Analysis)
Experience
-
2021 - Present Performance Manager - Commercial Team
Foodpanda Hong Kong - Implement LLM algorithm for competitor and product level analysis
- OpenAI API
- Develop dashboards, visualizations and regular reports to facilitate efficient data analysis on the topics of market competition, profit and loss, vendors acquisition and churn
- Tableau, Looker Studio, BigQuery
- Manage ETL process for data warehouse
- DBT, Airflow
- Build vendor grading machine learning model to predict vendor performance and churn risk
- Python, Pandas, NumPy, Scikit-learn, SciPy, Matplotlib, Seaborn, Plotly, Dash, Streamlit
- Implement LLM algorithm for competitor and product level analysis
-
2020 - 2021 Data Science Analyst
Hong Kong Telecom - Build machine learning model for customer churn prediction of LTS service and automate feature engineering and model training pipeline
- Create and schedule pipelines of web-crawling to extract data from external websites for sales strategies enhancement, competitor alert and machine learning model feature enrichment
- Conduct customer segmentation for Club Shopping from cross business unit data to understand customer interests and preferences for target marketing
Projects
-
Whitebox Transformer Implementation
- Implement CRATE architecture for image classification task on Fashion-MNIST dataset
- Visualize the low-rankness of the representation, autocorrelation of subspace learnt in each layer of MSSA
- Visualize the semantic meaning of each attention head
-
Resemblance of Cross Attention like Operator with Conditional GMM Denoiser
- Train conditional GMM denoise.
- Visualize the conditional sampling results and interpret the results
- Illustrate the resemblance of cross attention like operator with conditional GMM denoiser
-
(WIP) Mini World Model with CRATE
- Implement LeWM from scratch
- Replace ViT with a more interpretable architecture, CRATE, to study how spatiotemporal structure of data emerge in latent space
- Explore, with CRATE replacing ViT, whether attention patterns can enable unsupervised video segmentation
- Familiarize myself with Hydra and wandb for efficient experiment tracking