DSA Thrust Seminar: Towards Automated and Efficient Machine Learning Systems – An Evolutionary Multi-Objective Approach

With deep neural networks (DNNs) being the backbone, machine learning (ML) systems have seen widespread growth in applications recently, driven by both improvements in computing power and a massive amount of data. As DNN models become increasingly more complex, it’s well recognized now that manually designing ML systems is a computationally impractical endeavor. Automated machine learning (AutoML), on the other hand, presents a promising path to alleviate this painstaking process by posing the design of DNN as an optimization problem. However, existing AutoML approaches are primarily designed for standard benchmarks and do not readily scale to real-world scenarios, where there may not exist a sufficient amount of data, a proper set of training hyperparameters, differentiable objectives, etc. Meanwhile, evolutionary computation is a class of heuristic-based search techniques particularly suited for complex optimization problems, given its multi-point- and non-gradient-based nature. In this talk, Dr. Zhichao Lu will demonstrate how the intersection of evolutionary computation and deep learning can contribute to the aim of automating the design of efficient ML systems.