Yuan Ma

Skilled in developing and utilizing LLMs, fine-tuning with LoRA/DoRA, and implementing practical applications. Conducted research and applied Machine Learning and Deep Learning techniques in conversational recommender systems. Solely wrote crawlers (Scrapy) to collect the experimental dataset, fine-tuned various modern LLMs (Llama3, BERT, DialogFlow) for intent classification, developed a chatbot (REACT, Express, Flask) for research purposes, and deployed it in AWS (EC2, S3).

Developed a compressed 3D object detection model (Deep Learning) by introducing similarity and hybrid pruning techniques. This approach successfully reduced the model's size by 70% and eliminated 41% of the trainable parameters, while maintaining performance comparable to the original model. This innovation led to the issuance of a VW patent.

Locate top-tier open-source models for 3D object detection and multi-sensor fusion model (Deep Learning), then implement and assess them locally.

Explore the fields of Uncertainty in Deep Learning, Robust Learning, and Generative Adversarial Networks (GAN). Acquire knowledge in these domains and independently reconstruct the neural network from a paper for industrial testing and comparison in automatic optical inspection.

Apply Deep Learning to fake news detection, argument mining, and information nutrition labeling, and initially develop a big data platform (Spark). Finally published a paper as co-author.

Fine-tune pre-trained LLMs (Llama3, BERT, DialogFlow) for intent classification. Developed a chatbot (full-stack) equipped with intent recognition (NLP), interaction prompts, and product recommendation (RS) capabilities, deployed in AWS. Proposed meta-intents theory that blends high-level dialogue preferences with traditional feature-related preferences in Conversational Recommender Systems (CRS). Published seven papers as the first author in top-tier conferences and journals.

GPA:1.6 Argument Mining project: Refined the Deep Learning architecture and Feature Engineering, enhancing the accuracy of argumentative sentence detection (NLP) from 66% to 75%. Virtual Reality (VR) project: Solely managed all aspects of the VR software development using Unity 3D and JavaScript.