Detail

Automating Deep Learning Design using Neural Architecture Search

Deep Learning has had tremendous success in multiple tasks, mainly due to human expertise and ingenuity. Over the years, novel and optimized methods have been proposed, which incrementally improve the efficiency of the Deep Learning algorithms. However, designing optimal models, and applying existing ones to new tasks is extremely hard, requiring tremendous human effort, expertise and trial and error. This is a clear barrier to the mass application of Deep Learning solutions, such as Convolutional Neural Networks. Thus, Neural Architecture Search (NAS) emerged as a logical solution, where the goal is to automatically design optimal networks for a given problem, by automating the design and evaluation processes. NAS-designed networks have surpassed human-design networks, and shown to be efficient in multiple tasks. However, NAS methods still have drawbacks, such as human biases introduced during the development of the methods, and huge computational costs. In this talk, we will introduce NAS concepts, provide an introduction to the field, and show how NAS methods can still be improved, and leveraged in combination with human-crafted networks.