This paper introduces NASNet, a scalable image recognition architecture that leverages neural architecture search (NAS) to learn transferable convolutional cells. The key idea is to search for a convolutional cell on a smaller dataset (CIFAR-10) and then transfer it to a larger dataset (ImageNet). The NASNet search space is designed to enable transferability, allowing the same cell structure to be used across different input sizes and depths. The best cell found on CIFAR-10 is then stacked to form a NASNet architecture for ImageNet, achieving state-of-the-art results. The NASNet architecture is trained using a controller RNN that samples child networks and updates its weights via policy gradient. A new regularization technique, ScheduledDropPath, is introduced to improve generalization. NASNet achieves 2.4% error rate on CIFAR-10 and 82.7% top-1 accuracy on ImageNet, outperforming human-invented models with 9 billion fewer FLOPS. NASNet also performs well on object detection tasks, achieving 43.1% mAP on COCO with the Faster-RCNN framework. The NASNet search space is designed to be flexible, allowing for different computational demands by varying the number of convolutional cells and filters. This enables the creation of models with varying sizes, including a small version that achieves 74% top-1 accuracy on ImageNet, surpassing previous mobile-optimized models. The image features learned by NASNet are transferable to other computer vision tasks, demonstrating the model's versatility. The paper compares NASNet with other architectures, showing that it outperforms existing models in terms of accuracy and efficiency. The NASNet search space is designed to decouple architecture complexity from network depth, enabling efficient transfer to different datasets. The results show that NASNet achieves state-of-the-art performance on both CIFAR-10 and ImageNet with less computational demand than human-designed models. The paper also discusses the efficiency of different search methods, finding that reinforcement learning outperforms random search in this context.This paper introduces NASNet, a scalable image recognition architecture that leverages neural architecture search (NAS) to learn transferable convolutional cells. The key idea is to search for a convolutional cell on a smaller dataset (CIFAR-10) and then transfer it to a larger dataset (ImageNet). The NASNet search space is designed to enable transferability, allowing the same cell structure to be used across different input sizes and depths. The best cell found on CIFAR-10 is then stacked to form a NASNet architecture for ImageNet, achieving state-of-the-art results. The NASNet architecture is trained using a controller RNN that samples child networks and updates its weights via policy gradient. A new regularization technique, ScheduledDropPath, is introduced to improve generalization. NASNet achieves 2.4% error rate on CIFAR-10 and 82.7% top-1 accuracy on ImageNet, outperforming human-invented models with 9 billion fewer FLOPS. NASNet also performs well on object detection tasks, achieving 43.1% mAP on COCO with the Faster-RCNN framework. The NASNet search space is designed to be flexible, allowing for different computational demands by varying the number of convolutional cells and filters. This enables the creation of models with varying sizes, including a small version that achieves 74% top-1 accuracy on ImageNet, surpassing previous mobile-optimized models. The image features learned by NASNet are transferable to other computer vision tasks, demonstrating the model's versatility. The paper compares NASNet with other architectures, showing that it outperforms existing models in terms of accuracy and efficiency. The NASNet search space is designed to decouple architecture complexity from network depth, enabling efficient transfer to different datasets. The results show that NASNet achieves state-of-the-art performance on both CIFAR-10 and ImageNet with less computational demand than human-designed models. The paper also discusses the efficiency of different search methods, finding that reinforcement learning outperforms random search in this context.