How do CNNs form effective receptive fields and why is it different from theoretical receptive field?

The theoretical receptive field is the maximum input region influencing a neuron, while the effective receptive field is the smaller, Gaussian-like region that actually contributes most.

In CNNs, each layer increases the theoretical receptive field based on kernel size, stride, and depth. However, in practice, not all pixels in this region contribute equally. Due to weight initialization and gradient dynamics, the center pixels contribute more, forming a Gaussian-like distribution called the effective receptive field (ERF). This explains why very deep networks still rely heavily on central spatial information.