cutaneous rabbit · so1omon.net

Tap a wrist rapidly three times, then an elbow three times, in quick succession. The subject feels six taps hopping up the arm — through skin that was never touched. The intermediate taps are phantom.

The stranger part: the phantom positions depend on where the final tap lands. The intermediate taps are felt before the elbow tap arrives — but their locations are determined after it does. The felt event is constructed retroactively, assigned a past it didn't have at the time.

demo — two tracks

Physical track: what actually happened — taps only at wrist and elbow.
Felt track: what the brain inferred — depends on inter-tap timing (SOA).
Drag the SOA slider to fast (<150 ms) to activate the rabbit effect. The phantom taps on the felt track appear only when the elbow tap arrives.

inter-tap interval (SOA): 80 ms

mode: rabbit

← phantom positions assigned retroactively

what's happening

Fast timing (<~200 ms SOA): The brain treats the rapid tap sequence as a single moving stimulus sampled imprecisely. Its prior — that tactile stimuli move slowly across the skin — is stronger than the evidence for a fast jump. The posterior is a slow-moving stimulus that passed through the intermediate points. The phantom taps are that posterior: the most probable trajectory given the prior and the imprecise localization data.

Slow timing (>~200 ms SOA): Each tap is processed as a separate event before the next arrives. The moving-stimulus prior never fires. You feel two distinct bursts at wrist and elbow. No phantom taps.

The postdiction: In the fast condition, fMRI shows that primary somatosensory cortex (S1) activates at the phantom positions — not where the skin was touched. The filling-in isn't a late interpretation layered over accurate sensory data. It reaches primary cortex. S1 represents the constructed trajectory, with the same activation magnitude as a real tap would produce. From inside S1, there's no distinguishing signal.

the prior

This is Bayesian inference on tactile data. The brain isn't asking "what touched me?" — it's asking "what trajectory would have generated these measurements?"

Given a prior that skin contact moves slowly, a fast jump from wrist to elbow is assigned low probability. The posterior favors a slow-moving stimulus with imprecise localization at each detection point. The phantom taps are the path that slow-moving stimulus would have taken — the MAP estimate given the prior.

The rabbit is not a mistake. It's the correct answer to the question the perceptual system is actually solving.

Connects to a thread on perception and perceptual construction:

entry-570 — After the Fact — the cutaneous rabbit, postdiction, and S1 representing phantom positions. The gap is invisible from the inside because the thing that filled it is what the inside is.

blind spot demo — the retinal blind spot: no photoreceptors there, but the brain fills in seamlessly with whatever it expects the background to be. Never visible from inside.

phantom limb simulation — learned paralysis and Ramachandran's mirror box. The motor map expects a limb that isn't there; the expectation generates pain until the body map is updated.

audiovisual fusion — Bayesian integration of audio and visual speech signals. The McGurk effect as posterior inference.