Skip to content

Efference-Copy Cancellation

Predictive cancellation method — instantiates Predictive Residual Processing

Feed a copy of your own outgoing command into a forward model, predict the sensations you're about to cause, and subtract them — so only the world's part of the signal survives.

Much of what a system senses is the consequence of its own actions — noise it makes, motion it causes — and paying attention to that self-caused input is waste, or worse, confusion. Efference-Copy Cancellation takes a copy of each outgoing command, runs it through a forward model that predicts the sensory consequence that command will produce, and subtracts that prediction from the incoming signal, propagating only the unpredicted remainder. Its distinguishing feature is the source of the prediction: not the signal's past, not a statistical model of the world, but the actor's own motor/command copy. That is what lets it cleanly separate "I caused that" from "the world did that" — cancelling self-generated input specifically.

Example

A mobile robot's microphones pick up the room and the whirr of its own actuators; the self-noise can swamp the quiet external sounds it needs to hear. So a copy of every motor command is fed to a forward model that predicts the sound those actuators will make, and that prediction is subtracted from the microphone input, leaving the room's acoustics behind. The biological anchor is the reafference principle: the same trick an animal's nervous system uses to cancel the visual motion caused by its own eye movements — and the reason you can't tickle yourself, because the predicted self-touch is cancelled before it registers as surprising. Self-caused signal is nulled at the source; only exafference — the externally caused remainder — flows on.

How it works

  • Tap an efference copy — duplicate the outgoing command as it's issued.
  • Predict the consequence — run that copy through a forward model of the expected sensory return.
  • Compare and subtract — the comparator nulls the predicted self-signal against the actual afferent signal.
  • Propagate the residual — only the externally caused remainder (exafference) survives to downstream attention.

Tuning parameters

  • Forward-model fidelity — how accurately self-consequences are predicted; a crude model leaves self-signal uncancelled.
  • Timing alignment — the command copy must arrive matched in latency to the sensory return, or cancellation smears.
  • Cancellation aggressiveness — how completely the predicted component is subtracted; over-cancel and real external signal is deleted with it.
  • Adaptation / recalibration rate — how fast the forward model re-learns as the body or plant changes.

When it helps, and when it misleads

Its strength is cleanly removing the large, predictable, self-caused component of a signal, freeing attention and bandwidth for genuinely external events — the surprises that matter.

Its failure mode is a wrong or mistimed forward model: predict too little and self-signal leaks through; predict too much, or too late, and real external events get subtracted away as if you had caused them.[1] A changing plant that outruns recalibration drifts the model out of true. The classic misuse is cranking cancellation to silence self-noise and unknowingly erasing quiet external signals along with it. The discipline: keep the forward model calibrated and its timing matched, and monitor specifically for over-cancellation of real input.

How it implements the components

  • predictive_feedforward_model — the forward model that turns a command copy into an expected sensory consequence; the mechanism's defining organ.
  • prediction_comparator — the subtraction step that nulls predicted self-signal against the actual, leaving the residual.

It does not encode or transmit that residual efficiently (that's Delta or Differential Encoding), maintain a belief state with calibrated uncertainty (that's Bayesian Model Update), or model the expected patterns of external input (that's Anomaly Detection Model).

Notes

Compare Delta or Differential Encoding: both predict, then keep only the residual, but the prediction's source differs. Efference copy predicts from the system's own action (cancelling self-caused input); delta encoding predicts from the signal's past or a model (shrinking payload). The shared shape — predict, subtract, propagate the remainder — is the archetype; the source of the prediction is what makes these two distinct mechanisms.

References

[1] The reafference principle (von Holst & Mittelstaedt) and the related notion of corollary discharge describe exactly this: a copy of the motor command predicts and cancels self-generated sensation, so only externally caused exafference remains. It is the canonical natural instance of predictive residual processing. ↩