OpenPOWER Summit North America 2019

Deep learning has drawn growing attention in recent years. The impressive performance comes at scale of computational power, which makes it hard to execute especially on mobile devices. Besides, designing a good and efficient neural-net requires a lot of engineering efforts, let alone the following platform/hardware specific deployment. To reduce the human-bandwidth, we propose ProxylessNAS -- an efficient framework that automatically specializes neural architecture for different hardwares. With >74.5% top-1 accuracy, the measured latency of ProxylessNAS is 1.8x faster than MobileNet-v2, a widely used human design for mobile vision.