Local averaging smooths an image by replacing each pixel with the average of neighboring pixels. A convolution kernel performs this operation, washing out small noise and softening sharp edges before downsampling. section 3.3
For anti-aliased pooling, first compute max pooling with stride 1, then blur the result with a convolution using the desired stride. This mixes adjacent locations and reduces aliasing while retaining max-pooling behavior. section 14.1.2
A 3×3 image blur
Use a kernel whose nine entries are 1/9, convolve it with an MNIST digit image, and display the softened result. A binomial filter can instead concentrate more weight at the current location for anti-aliased pooling. section 14.1.2
