AlexNet was the first convolutional network to perform well on ImageNet, which is a 1000-class image classification task. It consists of eight hidden layers with ReLU activation functions, of which the first five are convolutional and the rest are fully connected.
The network starts by downsampling the input using an kernel with a stride of four to create 96 channels. It then downsamples again using a max pooling layer before applying a kernel to create 256 channels. There are three more convolutional layers with kernel size , eventually resulting in a representation with 256 channels. A final max-pooling layer yields a representation with 256 channels which is resized into a vector of length 9216 and passed through three fully connected layers containing 4096, 4096, and 1000 hidden units, respectively. The last layer is passed through the softmax function to output a probability distribution over the 1000 classes. The complete network contains 60 million parameters, most of which are in the fully connected layers.