In a convolutional neural network, what does "stride" refer to?

In a convolutional neural network, "stride" specifically refers to the distance that the filter moves across the input image as it scans for features. When the stride is set to a value greater than one, the filter will skip pixels as it passes over the image, leading to a reduced output size compared to using a stride of one, which allows the filter to evaluate every pixel in sequence. This parameter is crucial in determining the resolution and size of the feature maps produced by the convolutional layers, affecting both the performance and efficiency of the model.

Understanding stride is essential for optimizing the architecture and understanding how different strides influence the output dimensions and potentially the ability to capture spatial hierarchies in data. For instance, a larger stride can lead to down-sampling of the feature maps, which may enhance computational efficiency but could also result in loss of fine-grained features.