MnasNet: Platform-Aware Neural Architecture Search for Mobile

MnasNet addresses the difficulty of manually designing mobile CNNs that are simultaneously small, fast, and accurate given the huge space of architectural possibilities. The authors propose an automated mobile neural architecture search (MNAS) that explicitly incorporates model latency into the main objective, so the search identifies models with a good accuracy-latency trade-off. Unlike prior work that used inaccurate proxies such as FLOPS, MnasNet directly measures real-world inference latency by executing candidate models on mobile phones, and it introduces a novel factorized hierarchical search space that encourages layer diversity throughout the network.

Experiments show the approach consistently outperforms state-of-the-art mobile CNN models across multiple vision tasks. On ImageNet classification, MnasNet achieves 75.2% top-1 accuracy with 78ms latency on a Pixel phone, making it 1.8x faster than MobileNetV2 with 0.5% higher accuracy and 2.3x faster than NASNet with 1.2% higher accuracy. It also achieves better mAP for COCO object detection than MobileNets, demonstrating that platform-aware, latency-driven search produces broadly stronger mobile models.