Convolution layer
Slide small kernel over input. Kernel weights shared across positions → translation equivariance + parameter efficiency.
Advertisement
Pooling
Downsample via max or average over regions. Reduces spatial size + adds slight invariance. Modern architectures use strided conv instead.
Advertisement
Classic architectures
LeNet (1998, MNIST). AlexNet (2012, ImageNet breakthrough). VGG. ResNet (skip connections). EfficientNet. ConvNeXt.
Receptive field
Grows with depth. Dilated convolutions (Atrous) enlarge without increasing parameters — semantic segmentation.