ProxylessNAS is a novel approach that directly learns neural network architectures for large-scale tasks and target hardware platforms without the need for proxy tasks. It addresses the high computational and memory demands of conventional NAS algorithms by formulating the architecture search as a path-level pruning process, where an over-parameterized network containing all candidate paths is trained. The redundant paths are pruned at the end of training to obtain a compact optimized architecture. To reduce GPU memory consumption, the architecture parameters are binarized, allowing only one path to be active at runtime. Additionally, ProxylessNAS incorporates hardware metrics, such as latency, into the optimization objectives, enabling the learning of specialized neural architectures for different hardware platforms. Experiments on CIFAR-10 and ImageNet demonstrate the effectiveness of ProxylessNAS, achieving state-of-the-art performance with significantly fewer parameters and improved accuracy under latency constraints.ProxylessNAS is a novel approach that directly learns neural network architectures for large-scale tasks and target hardware platforms without the need for proxy tasks. It addresses the high computational and memory demands of conventional NAS algorithms by formulating the architecture search as a path-level pruning process, where an over-parameterized network containing all candidate paths is trained. The redundant paths are pruned at the end of training to obtain a compact optimized architecture. To reduce GPU memory consumption, the architecture parameters are binarized, allowing only one path to be active at runtime. Additionally, ProxylessNAS incorporates hardware metrics, such as latency, into the optimization objectives, enabling the learning of specialized neural architectures for different hardware platforms. Experiments on CIFAR-10 and ImageNet demonstrate the effectiveness of ProxylessNAS, achieving state-of-the-art performance with significantly fewer parameters and improved accuracy under latency constraints.