Unified Neural Architecture Search Framework for Computer Vision
ABOUT THIS PROJECT
At a glance
Many state-of-the-art algorithms in computer vision tasks such as classification, object detection, semantic segmentation, etc. are based on Deep Neural Networks (DNNs). In order to deploy these DNNs in autonomous driving settings, where these tasks are critical, highly efficient DNN architectures are needed. These architectures have traditionally been developed by hand, a very time-consuming process, and often ignore the constraints of deployment hardware. This leads to long design times, a lack of scalability in developing large numbers of new architectures, and lower network performance (both in efficiency and in accuracy metrics). In this project, we aim to address the above problems, by proposing an efficient, multi-task NAS framework capable of leveraging very large search spaces to design accurate, efficient models. We commit to open-source this framework through the BDD repository at the end of this project-year. This project presumes a fixed hardware target, and it is complementary to proposed work by Prof. Sophia Shao that will also include Neural Net accelerator search.
neural architecture search (NAS), multi-task NAS, efficient deep neural networks, expanded deep neural network search spaces