摘要
目的/意义梳理分析ChatGPT训练流程与理论模型,为医学研究提供参考借鉴。方法/过程系统梳理2018年以来GPT-1发布至今相关模型和流程文献资料,分析ChatGPT核心流程、理论模型及其创新点。根据现有资料分析ChatGPT预训练监督、自动评估以及强化学习近端策略优化(proximal policy optimization,PPO)模型3个层次技术构成。结合医学研究需求,分析人工智能技术面向医学信息领域应用的优化方向。结果/结论ChatGPT技术应用的突破是流程、算法和模型有效组合和不断迭代累积的结果,其模型及研究方法可以应用于医学文献自动化阅读与知识提取、基因研究与疾病风险评估等方面。
Purpose/Significance To analyze and outline the training process and theoretical model of ChatGPT,and to provide references for medical research.Method/Process The literatures on relevant models and process of GPT-1 released since 2018 are systematically reviewed.The core process,theoretical model,and innovative aspects of ChatGPT are analyzed.The three-level technical components of ChatGPT are examined,which include pre-training supervision,automatic evaluation,and proximal policy optimization(PPO)for reinforcement learning.Combining with the needs of medical research,the optimization direction of artificial intelligence(AI)technology for medical information application is analyzed.Result/Conclusion The breakthroughs in the application of ChatGPT technology result from effective combinations of processes,algorithms,and models through continuous iteration and accumulation.The models and research methods of ChatGPT can be applied in automated reading and knowledge extraction of medical literature,gene and disease risk assessment and so on.
作者
陈凌云
姚宽达
王茜
方安
李刚
CHEN Lingyun;YAO Kuanda;WANG Qian;FANG An;LI Gang(Institute of Medical Information,Chinese Academy of Medical Sciences&Peking Union Medical College,Beijing 100020,China;Agricultural Bank of China,Beijing 100005,China)
出处
《医学信息学杂志》
CAS
2023年第7期18-23,29,共7页
Journal of Medical Informatics
基金
国家重点研发计划项目子任务(项目编号:2022YFF0711902-2)
国家社会科学基金项目(项目编号:21CTQ016)
中国医学科学院医学与健康科技创新工程(项目编号:2021-I2M-1-056).
关键词
ChatGPT
预训练监督模型
强化学习近端策略优化模型
ChatGPT
pre-training supervised model
reinforcement learning proximal policy optimization(PPO)model
