Cutting edge neural architecture optimization in practice

Designing deep neural network architectures for different computer vision problems and different hardware setups is a daunting task even for teams of experts. Getting a well performing architecture and training strategy involves many iterations of manual tuning based on experience and intuition, and one has no theoretical guarantees of achieving the best possible performance.
Recent automated neural architecture search methodologies claim to solve more and more of these neural architecture design tasks. In this talk I would like to explain how can we apply deep learning not only to solve computer vision problems, but to create the neural network architectures themselves. I will give a brief overview of existing neural architecture search methods and interesting research directions that can automate the design process.

Deák-Meszlényi Regina 
Deep Learning for Vision Team Lead, Continental Automotive Hungary, Deep Learning Competence Center

Regina is a physicist and she obtained her PhD in medical image analysis, specifically machine and deep learning applications for cognitive neuroscience research at the Brain Imaging Centre of the Hungarian Academy of Sciences. She joined Continental’s Deep learning competence center for autonomous driving in 2018 as a deep learning expert for computer vision and neural architecture search and in her current position Regina provides technical supervision for all camera related methodology developments as team lead.