On November 2nd, Researcher Sun Yanan from Sichuan University was invited by our university to give an academic presentation. The presentation was conducted online and offline, and was hosted by Professor Mao Qirong, Dean of the School of Computer Science. More than 100 teachers and students from related fields participated in the presentation.
Researcher Sun mainly shared with everyone a performance predictor that can accelerate NAS algorithms. At present, many domestic and foreign scholars are committed to the research of performance predictors, aiming to save computational resources while preserving the accuracy of NAS prediction. This process mainly involves three aspects: search space, secondary search strategies, and search strategy evaluation. Teacher Sun gave a detailed introduction to the four major search spaces, as well as the search strategies involved in the search process, such as reinforcement learning, evolutionary computation, and gradients. He also conducted a detailed strategy evaluation of the mentioned strategies. Afterwards, we presented in detail the two algorithms currently being studied by our team, which are dedicated to implementing performance predictors and have achieved considerable success. However, these two algorithms still have significant shortcomings in practical applications, such as the narrow search space, the temporary inability to be used as universal predictors, and the issue of prediction accuracy. Finally, Teacher Sun pointed out that there are still many problems waiting to be studied in the performance predictor for accelerating neural architecture search, and encouraged everyone to actively engage in the exploration of performance accelerators.
After the lecture, Researcher Sun had sufficient communication and discussion with the teachers and students of our college, and patiently and meticulously answered the questions raised.