Kernel

A Kernel (also called a filter), in a Convolutional Neural Network, is a small matrix of learnable weights that slides across the input, computing a weighted sum at each position to produce a feature map.

For example, a 3x3 kernel looks at a 3x3 patch of pixels at a time. Kernels in early layers tend to learn simple features like edges, and deeper layers combine them into more complex patterns.

See Convolutional Layer and Pooling Layer.