• - Design occupancy detection system for residential building based on machine learning models.

• - Implement and fine-tune CNNs (Resnet50, FasterRCNN, YOLOv3) based on TensorFlow to identify and classify the collected image objects.

• - Design and implement four machine learning models (F1-score: Random Forest (0.98), Decision tree (0.93), SVM (0.87), KNN (0.93)) based on Scikit-learn to predict the occupancy status of residential buildings.

• - Implemented and fine-tuned the Bidirectional Encoder Representations from Transformers (BERT) model for document classification, leveraging its bidirectional contextual understanding for improved feature representation, resulting in a 11% increase in classification accuracy and ensuring the model's effectiveness in the target context.

• - Integrated the Generative Pre-trained Transformer (GPT) model into the Q/A chatbot architecture, leveraging its pre-trained capabilities to generate contextually relevant and accurate responses (question understanding and context-aware responses).

• - Build and deploy machine learning models and deep learning models in AWS/GCP, as well as dedicate to code maintenance and updates in open-source repositories (GitHub).

• - Develop and release via CI/CD and agile methodologies.

Project: Deep reinforcement learning model for user trading in double auction market

• - Propose an advanced deep reinforcement learning (DRL) model based on Deep Deterministic Policy Gradient to help users optimize their behaviors and obtain economic benefits in the market transactions.

• - Design neural networks by PyTorch to handle large number of high dimensional and continuous input data.

• - Fine-tune deep neural network based multi-agent reinforcement learning model (MARL), achieving an economic reward improvement of 24% to 49% compared to existing models.

• - Evaluate the performance of models and analyze feature importance to identify top factors.

• - Contributed to the design and development of user/item collaborative filtering algorithms and content filtering algorithm (two tower model) for recommendation system.

• - Conduct A/B testing and analysis to evaluate the performance of the recommendation system.

• - Collaborate with backend developers to integrate the recommendation system into the production environment.

• - Design privacy preserving models and Quantitative risk assessment models to de-identify joined data (including k-anonymity, Hashing, AES) from a variety of nonstandard data sources and/or from a set of documents only.

• - Design and develop efficient machine learning and deep learning models for large-scale data processing, detection system, recommendation system and modeling of user behaviors.

• - Develop attack models (blackhole attack, selective forward attack) and defense models (trust based defense model) in wireless network.

• - Conduct experiments to evaluate model performance, including designing experiments to evaluate the performance of models, and then analyze and interpret the results to improve model performance.

• - Investigate and debug issues in designed systems and provide timely resolutions.

• - Conduct codebase review, identifying areas for improvement and performing code refactoring to enhance cleanliness and maintainability.

Project: Representation learning for text-based documents (NLP)

• - Preprocess the dataset: clean the dataset and visualize the statistics of the dataset; build vocabulary based on TF-IDF.

• - Train two models (LDA model and Doc2Vec model by Gensim) upon the vocabulary; visualize topics, topic distribution (as features) for each document, learned word and document embedding space by t-SNE and LdaVis.

• - Conduct document clustering by K-means with Topics distribution and Doc2Vec representations; analyze their performance by Normalized Mutual Information (NMI) score.

Project: Network Security: Selective forwarding attacks and defense scheme in wireless network

• - Design an advanced selective forwarding attack model, which can flexibly select the targeted victim node and the percentage of forwarding packets with low detection rate (0.2-0.5).

• - Develop a lightweight trust-based defense solution to detect and eliminate malicious selective forwarding and blackhole nodes from the network, achieving detection rate (0.8-0.95).

Explain the experimental operating rules, teach experimental courses, grade assignments and guide students to complete the final circuit and PCB design.