Skip to content
ML Academy

Introduced on day 36

Receptive field

The set of input positions that one output unit depends on; it grows with kernel size, depth, dilation and stride.

In detail

For stride-1 layers it is 1+∑idi(ki−1)1 + \sum_i d_i (k_i - 1). Parameter sharing does not enlarge it: a shared filter still sees only its local patch. If the pattern you care about is longer than the receptive field (a 200-sample machine cycle with a 13-sample field), the model cannot use it unless pooling, stride, dilation or attention brings distant positions together.