报告题目:联邦环境下的多目标演化神经网络架构搜索
报告人:金耀初 院士、教授
报告时间:2022年10月24日下午16:00-17:00 (中国Beijing GMT +8.00)
报告地点:(线上)#腾讯会议:704-746-156
报告对象:相关学科教师、研究生等
主办单位:电气与信息工程学院,特种重载机器人安徽省重点实验室
报告人简介:
金耀初(Yaochu Jin),德国比勒菲尔德大学工程学院“洪堡人工智能教席教授”,兼任英国萨里大学计算机系“计算智能”讲席教授,欧洲科学院院士,IEEE Fellow。曾任芬兰国家创新局“Finland Distinguished Professor”、澳大利亚悉尼科技大学“杰出访问学者”。担任IEEE Transactions on Cognitive and Developmental Systems主编,Complex & Intelligent Systems共同主编。长期从事人工智能与系统科学的理论、算法和工程应用研究,特别是数据驱动的复杂系统进化优化、进化多目标机器学习、联邦学习与安全机器学习、演化发育系统与形态发育机器人学等。金耀初教授已发表学术论文500余篇,获美国、欧盟和日本专利9项。据Google Scholar,其论文被引用总次数34,000余次,h-index 92,入选Web of Science 2019-2021年度“全球高被引科学家”。
Abstract:
Neural architecture search is a useful step towards automated machine learning in that it is able todesign deep neural networks through a learning or optimization process. However, not much work has been done on search of optimal neural architectures when the training data are distributed on multiple sites and subject to privacy protection. This talk introduces two recent algorithms that perform neural architecture search in a federated environment using multi-objective evolutionary algorithms. We focus on the efficient representation of deep neural architectures, handling multiple objectives in learning, and reducing computational complexity of the neural architecture search process.