ProxylessNAS is a neural architecture search (NAS) method that directly learns architectures for large-scale tasks and hardware without using proxy tasks. It addresses the high GPU memory consumption issue of differentiable NAS by binarizing architecture parameters, reducing memory usage to the level of regular training. This allows for a large candidate set and avoids the restriction of repeating blocks. ProxylessNAS achieves strong results on CIFAR-10 and ImageNet, with a 2.08% test error on CIFAR-10 using 5.7M parameters, and a 3.1% higher top-1 accuracy than MobileNetV2 on ImageNet while being 1.2× faster. It also specializes neural architectures for hardware by directly optimizing for metrics like latency. ProxylessNAS reduces the computational cost (GPU hours and memory) to the same level as regular training, enabling efficient CNN design for different hardware platforms. The method uses a gradient-based approach to train binarized architecture parameters and handles non-differentiable hardware objectives through latency regularization loss or a REINFORCE-based algorithm. ProxylessNAS outperforms previous NAS methods in terms of accuracy and efficiency, achieving state-of-the-art results on CIFAR-10 and ImageNet under various hardware constraints. It also provides insights into efficient CNN design for different hardware architectures.ProxylessNAS is a neural architecture search (NAS) method that directly learns architectures for large-scale tasks and hardware without using proxy tasks. It addresses the high GPU memory consumption issue of differentiable NAS by binarizing architecture parameters, reducing memory usage to the level of regular training. This allows for a large candidate set and avoids the restriction of repeating blocks. ProxylessNAS achieves strong results on CIFAR-10 and ImageNet, with a 2.08% test error on CIFAR-10 using 5.7M parameters, and a 3.1% higher top-1 accuracy than MobileNetV2 on ImageNet while being 1.2× faster. It also specializes neural architectures for hardware by directly optimizing for metrics like latency. ProxylessNAS reduces the computational cost (GPU hours and memory) to the same level as regular training, enabling efficient CNN design for different hardware platforms. The method uses a gradient-based approach to train binarized architecture parameters and handles non-differentiable hardware objectives through latency regularization loss or a REINFORCE-based algorithm. ProxylessNAS outperforms previous NAS methods in terms of accuracy and efficiency, achieving state-of-the-art results on CIFAR-10 and ImageNet under various hardware constraints. It also provides insights into efficient CNN design for different hardware architectures.