1x1 Convolution In Detail - how 1x1 convolution layer works

Often, the layers in deep convolution networks have an increasing number of filters from the first layer to the last. Also, the kernel size goes down or stays the same in some models. In the early convolution networks like the AlexNet, the authors used huge filter sizes (11 11 kernels in the first layer of AlexNet). And they gradually reduced the kernel size for the layers deep in the network. This is due to the increased receptive field of the later layers on the input image. While having more filters and a huge number of layers helps the model capture more abstract features, it increases the model's size among other problems.