Stare at a waterfall, then look at the rocks — they drift upward. A neuron fed a steady input stops reporting it, and when the input vanishes it overshoots the other way: the aftereffect. Press and hold the screen to feed the neuron, slide how fast it tires — and watch the trace tire and rebound.
A leaky-integrator adaptation model: an internal state a chases the stimulus s with time constant τ (slide it), and the firing rate is the un-adapted part, r = gain·(s − a). Hold the screen to feed a steady stimulus and the trace spikes at onset, then decays as a catches up (the neuron 'stops seeing' the constant input); release and s drops to 0 while a is still high, so r goes negative — the aftereffect / negative afterimage. The dynamics are integrated live and the trace is real. Standard sensory-adaptation / firing-rate modelling.
A single first-order leaky integrator stands in for real adaptation, which has multiple time-scales, spike-rate saturation, and channel-specific mechanisms; firing rate here is a continuous number, not real spikes. The onset transient, the decay, and the negative aftereffect are exactly what this real model produces — the figure is that one time-constant captures a whole sense.