Hongchao(Cosimo) Fang

I am a graduate student from Northeastern University , majoring in Computer Science. Before that, I graduated from Central University of Finance and Economics， China. I am fortunate to work under the supervision of Dr. Min, Prof. Hannaneh in UW, Weiyan Shi from Stanford University, and Prof. Xie in UCSD.

I am especially grateful for Prof. Peng in the Chinese Academy of Sciences who brought me into the magical journey of research.

Scholar  /  Email  /  LinkedIn  /  Github  /  Curriculum Vitae

• Aug 28, 2023 - Glad to announce that I'll be joining CHATS-Lab this fall as a researcher, working on a brand-new topic of fine-grained control on interactive language models with Dr. Weiyan Shi from Stanford University!
• Jan 15, 2023 - Excited to share that I am selected to join H2Lab, UW as a research student, supervised by Dr. Min and Prof. Hannaneh
• Nov 28, 2022 - One new paper based on CERT was published by TACL
• Oct 12, 2022 - CERT gained over 200 citaions! So glad that this paper was liked by so many researchers.
• May 8, 2022 - I will join Amazon as a Software Engineer Intern. First time working in the industry, looking forward to that.
• Mar 16, 2020 - I will join AI4H Lab in UCSD, mentored by Prof. Xie
• Apr 9, 2019 - I will join the edge computing Lab in ICT, Chinese Academy of Science, mentored by Prof. Peng

I work on applying self-supervision on Large Language Models at an immense scale. I have done several works regarding the application of contrastive learning on text classification tasks, e.g., CERT. I'm currently working on applying self-supervised learning on generation tasks to explore the ability of large language models learning personality from domain dialogues.

We propose a four-level optimization framework that performs data augmentation and contrastive learning end-to-end, to enable the augmented data to be tailored to the contrastive learning task.

We propose CERT: Contrastive self-supervised Encoder Representations from Transformers, which pertrains language representation models using contrastive self-supervised learning at the sentence level

We build large-scale medical dialogue datasets – MedDialog, which is the largest medical dialogue dataset to date. We pre-train several dialogue generation models on the Chinese MedDialog dataset, including Transformer, GPT, BERT-GPT, and compare their performance. It is shown that models trained on MedDialog are able to generate clinically correct and human-like medical dialogues. We also study the transferability of models trained on MedDialog to low-resource medical dialogue generation tasks. It is shown that via transfer learning which finetunes the models pre-trained on MedDialog, the performance of medical dialogue generation tasks with small datasets can be greatly improved, as shown in human evaluation and automatic evaluation.

• My English name: Cosimo, comes from the protagonist in Calvino's novel: The Baron in the Trees. It tells the story of a baron who stayed on the trees all his life, maintaining his integrity and pursuing his ideals.
• I am a big rock fan with a wide interest in Punk, Shoe-gazing, and Jazz.
• I watch movies a lot, especially for French New Wave films. My favorite director is Jean-Luc Godard
• Besides movies and rock, I am also an enthusiast for extreme sports, including climbing, snowboarding, and skateboarding.

• Academic Excellence Award at Central University of Finance and Economics

Website design from Jon Barron