An adaptive dilation rate is a dynamic parameter in convolutional neural networks that adjusts the spacing between kernel elements according to input features, rather than remaining a fixed, global hyperparameter. In conventional dilated convolution, a static dilation rate expands the receptive field by a uniform factor across an entire layer, which can restrict the model when processing complex data with varying scales and spatial frequencies. By enabling the dilation rate to adapt dynamically across spatial locations, channels, or input instances, the network can tailor its receptive field size to local context. This mechanism allows the model to preserve fine details with denser sampling in high-detail areas while expanding the receptive field in uniform or background regions to capture broader contextual information, thereby improving representation learning across various visual tasks.
