Research Experience

Text Style Transfer Student Researcher

Massachusetts Institute of Technology

• Used Sequence Classification and SHAP for deletion, identifying and explaining style-related word distributions.
• Deconstructed Bert’s Masked model for the generation component and trained it on the target text style.
• Achieved the generation of target style vocabulary, replacing words that contributed to the original style.
• Integrated spacy for POS tagging and glove for semantic similarity, with attention to ensure a balance between text style and semantic retaining. Achieved a 10% efficiency improvement over SOTA models.

Multimodal Analysis Student Researcher

Massachusetts Institute of Technology

• Developed an emotion extraction model by combining image emotion retrieval and text emotion retrieval.
• Contained seven emotion labels and had 9% higher prediction accuracy especially on irony class.
• Constructed image caption generation model with Resnet50 to extract image attributes.
• Combined the image attributes, retrieved emotion information and original text comments with BERT.
• Employed self-attention in the final layer to achieve weighted integration of image and text information.

Generative Adversarial Network Student Researcher

Carnegie Mellon University

• Implemented ViTGAN and SAGAN as solutions for face spoofing, achieving facial restoration in low-light conditions. Investigated the efficacy of attention mechanisms in capturing facial details.
• Annotated a dataset distinguishing between dark and light settings, directing the discriminator to enhance the generator’s transition from dark to light facial generation.
• Determined that ViTGAN’s self-attention layer outperformed SAGAN in capturing facial nuances, leading to superior restoration in darker environments.

Project Experience

Software Development Intern

Neusoft Corporation

• Built a full-featured B2B e-commerce platform. Enabled functions such as shopping cart, order management, product management, product information maintenance and cargo transportation management.
• Utilized Vue for front-end, Redis for caching, and Elasticsearch for quick data retrieval.
• Realized Pass-through communication between different hosts with NAT.
• Utilized Springboot as the back end, SpringCloud to construct a distributed system with the interaction between subsystems, and SpringSecurity for declarative and secure access control.

Music Recommendation Platform Developer

Northeastern University

• Implemented machine learning and deep learning recommendation algorithms using Spark and Pytorch frameworks.
• Developed admin and user systems using Spring Boot and VUE frameworks.
• Utilized BERT models for learning music audio features and classifying music styles, addressing the cold start problem in music recommendations.
• Optimized music similarity matrices by using TF-IDF algorithm and music tags, enhancing recommendation accuracy.
• Generated user-rating-based collaborative filtering music similarity matrices using ALS algorithm and performed weighted blending.
• Utilized MySQL as the user business database, and MongoDB as the recommendation database.
• Achieved real-time dynamic recommendations based on user ratings through Kafka and Flume.