Prof. Yanan Sun

Sichuan university, China

Speech Title: Performance Predictor for Neural Architecture Search 

Abstract：

Neural architecture search (NAS) is the key technique to automate the design of deep neural networks, and is also the research focus among the artificial intelligence community recently. NAS requires a large amount of deep neural networks to be trained for obtaining accuracy, which in turn guides the search algorithm efficiently and effectively running. However, training a such number of deep neural networks requires intensive computation resources, especially the GPUs, which significantly challenges the development of NAS techniques. Performance predictors could directly evaluate the performance of deep neural networks without GPU resources, and has the great potential to greatly reduce the computation complexity of NAS. I will introduce two works of our team, which have been published in ICCV2021 and NeurIPS2022. These two works are novel algorithms to build high-performance predictors from the aspects of architecture augmentation and architecture domain transferring learning.

Biography:

Prof. Yanan Sun, researcher (Senior), doctoral supervisor. Selected candidates of the Tianfu Emei Talent Program by the Organization Department of Sichuan Provincial Party Committee, Sichuan University Double Hundred B talents, Sichuan Provincial Scientific and Technological Innovation talents, and Sichuan Provincial high-level overseas students have been included in the World's Top 2% Scientists 2021 list released by Stanford University.

In June 2017, he graduated from Sichuan University with a doctor's degree in engineering. From August 2015 to February 2017, he was assigned to Oklahoma State University. From July 2017 to March 2019, worked at Victoria University of Wellington, New Zealand; At the end of March 2019, he was introduced by Sichuan University as a special researcher from overseas, and later added as a doctoral supervisor by the Degree assessment Committee of Sichuan University and selected as the Double hundred B talents of Sichuan University (Green channel). His research interests include machine learning, evolutionary computing and its applications to neural architecture search. In terms of scientific research, he has published 40 high-level papers in the past five years, among which 16 papers have been published in the global Top 1% journals under the categories of "Computer Science and Artificial Intelligence", "Computer Science, Theory and Methods", "Computer Architecture" and CCF Class A international conference as the first/corresponding author. 7 papers were selected as the Top 1% "highly cited papers" and Top 0.1% "hot papers", and 2 papers were selected as the research frontier papers of IEEE Computational Intelligence Society.

In terms of teaching, it was awarded the 2020 Excellent Instructor of Top-notch Innovative Talents Training of Sichuan University, the Excellent instructor of Innovation and Entrepreneurship and Practical Education of College Students, and the Excellence Award of inquiry-based small-class teaching Quality. Guided undergraduate students to win several important awards in national competitions, presented research results and made invited oral presentations at several mainstream international conferences. In terms of postgraduate training, I supervised postgraduate students in CCF Class A conference and AI top Trans. He published his scientific research results as the first author and was reported on the homepage of the school's official website. He was awarded the National Scholarship for Graduate Students (directly recognized) and the first-class Scholarship for graduate students of Sichuan University. In terms of international cooperation, we have close cooperation with many famous research teams from Hong Kong, the United States, the United Kingdom, Germany, Australia, New Zealand and other countries. Both sides have stable joint training programs for graduate students and exchange mechanism for students. In terms of project cooperation, we have undertaken/completed 8 national and provincial vertical projects, and 3 horizontal projects from Baidu and Huawei.