Selective Kernel Networks (SKNets) are introduced to enable neurons in Convolutional Neural Networks (CNNs) to adaptively adjust their receptive field sizes based on input information. The proposed SK unit consists of three operations: Split, Fuse, and Select. Split generates multiple paths with different kernel sizes, Fuse aggregates information from these paths, and Select uses attention mechanisms to adaptively select the most relevant features. SKNets are computationally efficient and outperform existing state-of-the-art models on ImageNet and CIFAR benchmarks with lower model complexity. They demonstrate superior performance in object recognition by allowing neurons to dynamically adjust their receptive fields based on input stimuli. SKNets are also applicable to lightweight models like ShuffleNetV2. The method is validated through experiments on ImageNet and CIFAR datasets, showing that SKNets achieve significant improvements in accuracy and efficiency. The results indicate that the adaptive selection mechanism enables neurons to effectively capture multi-scale features, leading to better performance in object recognition tasks. The paper also discusses related work, including multi-branch convolutions, grouped/depthwise/dilated convolutions, and attention mechanisms, and highlights the advantages of the proposed SKNets in terms of efficiency and adaptability.Selective Kernel Networks (SKNets) are introduced to enable neurons in Convolutional Neural Networks (CNNs) to adaptively adjust their receptive field sizes based on input information. The proposed SK unit consists of three operations: Split, Fuse, and Select. Split generates multiple paths with different kernel sizes, Fuse aggregates information from these paths, and Select uses attention mechanisms to adaptively select the most relevant features. SKNets are computationally efficient and outperform existing state-of-the-art models on ImageNet and CIFAR benchmarks with lower model complexity. They demonstrate superior performance in object recognition by allowing neurons to dynamically adjust their receptive fields based on input stimuli. SKNets are also applicable to lightweight models like ShuffleNetV2. The method is validated through experiments on ImageNet and CIFAR datasets, showing that SKNets achieve significant improvements in accuracy and efficiency. The results indicate that the adaptive selection mechanism enables neurons to effectively capture multi-scale features, leading to better performance in object recognition tasks. The paper also discusses related work, including multi-branch convolutions, grouped/depthwise/dilated convolutions, and attention mechanisms, and highlights the advantages of the proposed SKNets in terms of efficiency and adaptability.