The paper introduces the Residual Attention Network (RAN), a convolutional neural network that incorporates an attention mechanism to enhance feature representation. The RAN is composed of multiple Attention Modules, each generating attention-aware features that adaptively change as the network depth increases. The attention modules use a bottom-up top-down feedforward structure to process features and generate attention masks, which are then used to control the learning process. The paper proposes attention residual learning to train very deep RANs, allowing for significant performance improvements without increasing computational complexity. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet datasets demonstrate that the RAN achieves state-of-the-art performance in object recognition tasks, outperforming ResNet-200 with a 0.6% top-1 accuracy improvement and reduced computational requirements. The RAN also shows robustness against noisy labels, making it a promising approach for image classification.The paper introduces the Residual Attention Network (RAN), a convolutional neural network that incorporates an attention mechanism to enhance feature representation. The RAN is composed of multiple Attention Modules, each generating attention-aware features that adaptively change as the network depth increases. The attention modules use a bottom-up top-down feedforward structure to process features and generate attention masks, which are then used to control the learning process. The paper proposes attention residual learning to train very deep RANs, allowing for significant performance improvements without increasing computational complexity. Extensive experiments on CIFAR-10, CIFAR-100, and ImageNet datasets demonstrate that the RAN achieves state-of-the-art performance in object recognition tasks, outperforming ResNet-200 with a 0.6% top-1 accuracy improvement and reduced computational requirements. The RAN also shows robustness against noisy labels, making it a promising approach for image classification.