Delta-Aug: Cross-Modal Hard Feature Mining for Few-Shot Learning
ACMLC '24: Proceedings of the 2024 6th Asia Conference on Machine Learning
and Computing
Link: Open Access Paper from ACM Digital
Library
- Researched LLMs and VLMs and discovered hard classes degrade few-shot performance.
- Proposed DeltaAug using CLIP for multi-modal feature expansion.
- Achieved 76.16% accuracy over 11 datasets, outperforming Wise-FT (71.64%).
Content Moderation with Multi-Modal LLM/VLM
Nov 2024 β Present | Advisor: Prof. Xiaowei Yang, Duke University
- Building system analyzing text, images, and video using LLM/VLMs.
- Designing full pipeline for online content moderation.
MQTT Protocol Research
Jul 2021 β Dec 2023 | Advisor: Prof. Fu Chen, CUFE
- Studied MQTT protocol and implemented it using Mosquitto reference.
- Explored QoS and large-scale codebase structure.
Stock Market Forum Analysis System
Jul 2023 β Nov 2023 | Advisor: Prof. Dongli Han, CUFE
- Developed full-stack site with Spring Boot and Vue.
- Designed review analysis with Sentence-BERT; used MySQL + MongoDB.
Convexity in Beta Distributions
Oct 2021 β Apr 2023 | Advisor: Prof. Jinting Wang, CUFE
- Analyzed convexity/log-convexity properties of Beta distribution.
- Submitted findings to Statistics & Probability Letters.
Lucene IR System & Website
Sep 2022 β Nov 2022 | Advisor: Prof. Yan Wang, CUFE
- Built full-stack IR system using Lucene and Grobid.
- Implemented fuzzy search, image indexing, PDF metadata extraction.