multimodal

Interest: image-to-text, text-to-image, image-text alignment; computer vision, natural language processing;- in collaboration with OpenAI; DALLE-3- Core team member of the Florence project, focusing on large-scale multimodal pretraining.- Primary contributor to the development of the GIT/GIT2 model, pretrained on 10B+ image-text pairs (world's largest), achieving state-of-the-art results on 14 vision-language benchmarks (image/video, caption/question-answering, OCR).- Conducted various research projects, such as MMReact, NUWA-Infinity, GIT, GRiT, UniTab, ReCo, UFO, MiniVLM, e.t.c.- Delivered image recognition and vision-language models to product.- Developed internal productivity tools, such as open-sourced act and azfuse.

Distributed storage system; TLA+; deployment

Algorithm development for fashion recognition, especially by deep learning.

Hashing-based nearest neighbor search and deep learning for image recognition.

Improve the accuracy of approximate nearest neighbor search based on hashing/quantization.

Collaborative web browsing by multiple devices, and binary encoding for nearest neighbor search.