Selective Kernel Networks

TLDR

Standard CNNs use fixed receptive field sizes, whereas neuroscience shows visual cortical receptive fields are stimulus‑dependent, a property rarely incorporated into CNN design. This work introduces a dynamic selection mechanism that lets each neuron adjust its receptive field size according to multiple input scales. The Selective Kernel (SK) unit fuses several branches with different kernel sizes using softmax‑based attention, producing variable effective receptive fields that are stacked to form SKNets. SKNets achieve state‑of‑the‑art performance on ImageNet and CIFAR with lower complexity, and analyses confirm that neurons adaptively capture objects at varying scales. Code and pretrained models are available at https://github.com/implus/SKNet.

Abstract

In standard Convolutional Neural Networks (CNNs), the receptive fields of artificial neurons in each layer are designed to share the same size. It is well-known in the neuroscience community that the receptive field size of visual cortical neurons are modulated by the stimulus, which has been rarely considered in constructing CNNs. We propose a dynamic selection mechanism in CNNs that allows each neuron to adaptively adjust its receptive field size based on multiple scales of input information. A building block called Selective Kernel (SK) unit is designed, in which multiple branches with different kernel sizes are fused using softmax attention that is guided by the information in these branches. Different attentions on these branches yield different sizes of the effective receptive fields of neurons in the fusion layer. Multiple SK units are stacked to a deep network termed Selective Kernel Networks (SKNets). On the ImageNet and CIFAR benchmarks, we empirically show that SKNet outperforms the existing state-of-the-art architectures with lower model complexity. Detailed analyses show that the neurons in SKNet can capture target objects with different scales, which verifies the capability of neurons for adaptively adjusting their receptive field sizes according to the input. The code and models are available at https://github.com/implus/SKNet.