Receptive fields in convolutional networks

WHAT THIS TESTS Whether you understand how spatial context accumulates through a network and can connect that to why architectures are designed with certain depths and strides.

A GOOD ANSWER COVERS The receptive field of a neuron is the region of the original input image that can affect its value. A first-layer three by three neuron sees three by three pixels. Stacking conv layers grows it additively: two three by three layers give a five by five field, three give seven by seven. Larger kernels enlarge it faster per layer. Stride and pooling are the strongest multipliers, because they downsample, so each step in a deeper layer corresponds to many more input pixels. Depth therefore expands the field both additively and, through strides, multiplicatively. This matters because a neuron can only respond to information inside its receptive field, so to recognize a large object or use global context, deep layers must have fields large enough to span it.

COMMON WRONG ANSWERS Thinking the receptive field grows only with kernel size and ignoring stride and depth. Assuming all pixels in the receptive field contribute equally; the effective receptive field is smaller and roughly Gaussian, with center pixels weighted far more. Forgetting dilation as a way to grow the field without extra parameters.

LIKELY FOLLOW-UPS What is the difference between theoretical and effective receptive field. How does dilated convolution help. Why do segmentation models care about large receptive fields. How would you increase the field without adding much compute.

ONE CONCRETE EXAMPLE In a network with several stride-two stages, a neuron deep in the last stage may have a theoretical receptive field covering most of a two-twenty-four by two-twenty-four image, enabling whole-object reasoning. But measuring its effective field shows only the central region strongly influences it, which is why detectors of large objects still benefit from extra context modules or dilation even when the theoretical field already spans the image.