directly into the architecture. These biases dramatically reduce parameter count while making networks equivariant to spatial translations. CNNs have dominated computer vision for over a decade, and their principles extend to audio, text, and scientific computing.