Divisive inhibition is a gain knob, not a threshold
Subtractive inhibition moves a curve. Divisive inhibition rescales it. On a sensor pixel array that difference is the difference between clipping and seeing.
A neuron receiving strong drive saturates. The standard fix in a network is to subtract: raise the threshold until the response falls back into range. That works for one input level and fails for every other one, because subtraction does not change the slope — it slides the whole curve to the right and throws away the small signals at the bottom.
Divisive inhibition scales instead. Feed the pooled activity of a neighbourhood into the denominator and the transfer function stretches to cover whatever range the input happens to occupy. In insect visual systems this is how contrast sensitivity survives a five-order-of-magnitude swing in ambient light. It is also cheap in analog hardware: a shared line, not a per-pixel calibration.
Drag the two controls below. Watch what happens to the small-signal region — the part of the curve below the marker — as you raise each kind of inhibition.
Response of a single unit as input level rises. Copper has no inhibition; teal divides by pooled activity; gold subtracts a constant. Divisive inhibition changes the slope, so the small-signal region below the cream marker stays informative.
Everything here is a toy: one static nonlinearity, no time, no noise, no adaptation, and a pooled inhibition term that stands in for an entire circuit. Treat the curves as intuition about slope and range, not as a model of any measured neuron or pixel.
The practical consequence for hardware is that the inhibitory signal wants to be an analog quantity shared across a tile, computed from the same charge domain as the drive. That is what Lateral Divisive Inhibition for Dynamic Gain Control does in a pixel array, and it is why the patent filing on it is about the array rather than the neuron.