Progressive Neural Architecture Search (PNAS) is a method for learning the structure of convolutional neural networks (CNNs) that is more efficient than recent state-of-the-art methods based on reinforcement learning (RL) and evolutionary algorithms. PNAS uses a sequential model-based optimization (SMBO) strategy, searching for structures in order of increasing complexity while learning a surrogate model to guide the search. This approach is up to 5 times more efficient than the RL method of Zoph et al. (2018) in terms of number of models evaluated and 8 times faster in terms of total compute. The structures discovered by PNAS achieve state-of-the-art classification accuracies on CIFAR-10 and ImageNet. The paper introduces a method that learns a CNN matching previous state-of-the-art accuracy while requiring 5 times fewer model evaluations during architecture search. It builds on the structured search space proposed by [41], where the search algorithm searches for a good convolutional "cell" rather than a full CNN. A cell contains B "blocks," each combining two inputs with an operation and then combining the results. This cell structure is stacked multiple times to form the final CNN. The modular design allows easy transfer between datasets. PNAS uses heuristic search to explore the space of cell structures, starting with simple models and progressing to complex ones. At each iteration, a set of candidate cells is trained and evaluated. A surrogate model predicts the performance of structures without training them, allowing the search to focus on promising candidates. This process is repeated until the desired number of blocks is reached. The paper compares PNAS with other methods, including random search and NAS. PNAS is more efficient, requiring fewer models to be evaluated and achieving better performance. It also shows that PNAS can find models with similar accuracy to NAS but using significantly less compute. The results on CIFAR-10 and ImageNet demonstrate that PNAS achieves state-of-the-art performance, with PNASNet-5 achieving higher accuracy than previous methods while using less compute. The method is also effective for ImageNet classification, showing strong correlation between CIFAR-10 and ImageNet accuracy.Progressive Neural Architecture Search (PNAS) is a method for learning the structure of convolutional neural networks (CNNs) that is more efficient than recent state-of-the-art methods based on reinforcement learning (RL) and evolutionary algorithms. PNAS uses a sequential model-based optimization (SMBO) strategy, searching for structures in order of increasing complexity while learning a surrogate model to guide the search. This approach is up to 5 times more efficient than the RL method of Zoph et al. (2018) in terms of number of models evaluated and 8 times faster in terms of total compute. The structures discovered by PNAS achieve state-of-the-art classification accuracies on CIFAR-10 and ImageNet. The paper introduces a method that learns a CNN matching previous state-of-the-art accuracy while requiring 5 times fewer model evaluations during architecture search. It builds on the structured search space proposed by [41], where the search algorithm searches for a good convolutional "cell" rather than a full CNN. A cell contains B "blocks," each combining two inputs with an operation and then combining the results. This cell structure is stacked multiple times to form the final CNN. The modular design allows easy transfer between datasets. PNAS uses heuristic search to explore the space of cell structures, starting with simple models and progressing to complex ones. At each iteration, a set of candidate cells is trained and evaluated. A surrogate model predicts the performance of structures without training them, allowing the search to focus on promising candidates. This process is repeated until the desired number of blocks is reached. The paper compares PNAS with other methods, including random search and NAS. PNAS is more efficient, requiring fewer models to be evaluated and achieving better performance. It also shows that PNAS can find models with similar accuracy to NAS but using significantly less compute. The results on CIFAR-10 and ImageNet demonstrate that PNAS achieves state-of-the-art performance, with PNASNet-5 achieving higher accuracy than previous methods while using less compute. The method is also effective for ImageNet classification, showing strong correlation between CIFAR-10 and ImageNet accuracy.