Learning to Automate Machine Learning Lifecycle

Wednesday, March 04, 2020
10:30 AM to 11:30 AM
EER 3.646
Free and open to the public

Machine Learning (ML) is automating the world, but to what extent can ML itself be automated? In this talk, I will present our innovative efforts towards the goal of automating the ML lifecycle (e.g., model design, optimization and deployment), by leveraging the data-driven power. The first part will focus on the automatic discovery and tuning of optimization algorithms from data. By learning to adapt their behaviors, those data-driven optimizers can reduce their complexities by orders of magnitude while improving the model accuracy, compared to their classical counterparts. Furthermore, our learned optimizers preserve favorable theoretical guarantees such as convergence and robustness. The next part will demonstrate how to automatically search for the best model architectures with the help of problem priors, validated with application examples from complicated real-world computer vision tasks. The talk will be concluded by discussing our ongoing works, including the automatic algorithm-hardware co-design for practical deployment, and eventually, the co-optimization of the entire ML lifecycle.


Zhangyang (Atlas) Wang

Zhangyang (Atlas) Wang

Texas A&M University

Dr. Zhangyang (Atlas) Wang has been an Assistant Professor of Computer Science and Engineering at Texas A&M University, since 2017. He was a Ph.D. student of Electrical and Computer Engineering, at the University of Illinois at Urbana-Champaign, working with Professor Thomas S. Huang, from 2012 to 2016. Dr. Wang is broadly interested in the fields of machine learning, computer vision, optimization, and their interdisciplinary applications. His latest interests focus on addressing automated machine learning (AutoML), learning-based optimization, and efficient deep learning. He has received many research awards and scholarships, including a TAMU Engineering Genesis Award and three research challenge prizes. His research is gratefully supported by NSF, DARPA, ARL/ARO, as well as a number of industry and university grants