Sensory Feedback

Abstract

Sensory feedback, which MeSH classifies under biofeedback, is the afferent information that returns to the nervous system as a consequence of the body's own movement. The reafference principle distinguishes this self-produced signal from externally caused stimulation by means of an efference copy of the outgoing motor command. Forward models use that copy to predict the sensory consequences of action, while optimal feedback control specifies when a returning signal should correct a movement in progress. Deafferentation, the loss of feedback, degrades coordination even when muscle strength is intact. Modern neuroprostheses now restore artificial feedback through intraneural and cortical stimulation.

Keywords: reafference, efference copy, forward model, proprioception, motor control

Sensory feedback is the stream of afferent signals that the nervous system receives as a result of its own motor output, and its central problem is that the same receptors are driven both by the world and by the mover (von Holst & Mittelstaedt, 1950). A retinal image shifts identically whether the eye turns or the scene moves; a muscle spindle fires whether a limb is displaced by its own contraction or by an external push. Resolving that ambiguity, using the returning signal to guide action without being fooled by it, is the function that ties perception to movement, and it organises this article from the reafference principle through forward models, the sensory channels that carry feedback, the evidence from patients who have lost it, and the prostheses now built to give it back.

What Sensory Feedback Is

Sensory feedback is afferent activity that arises because the organism moved. It is set against feedforward signals, the motor commands that leave the nervous system, and the two form a loop: a command produces movement, movement stimulates receptors, and the resulting afference returns to influence the next command (Miall & Wolpert, 1996). The loop is closed when the returning signal is actually used to adjust the ongoing command and open when the movement runs to completion before any feedback can act. Because neural conduction, mechanical delay and processing together impose latencies of the order of a hundred milliseconds, purely closed-loop control is stable only for slow movements; fast actions must be driven in part by prediction (Miall & Wolpert, 1996).

The signal that returns is not homogeneous. Von Holst and Mittelstaedt distinguished reafference, the stimulation caused by the animal's own action, from exafference, the stimulation caused by events in the world (von Holst & Mittelstaedt, 1950). Both reach the same sense organs, so the nervous system cannot separate them at the receptor. It separates them centrally, by keeping a copy of what it just commanded and comparing the prediction that copy affords against the afference that arrives.

The Reafference Principle

In 1950 two independent papers proposed the same solution to the reafference problem. Von Holst and Mittelstaedt, working on the optokinetic responses of insects and fish, argued that each motor command deposits an efference copy whose expected sensory consequence is held and then subtracted from the afference that returns; only the residue, the exafference, reaches perception (von Holst & Mittelstaedt, 1950). In the same year Sperry, studying fish whose eyes he had surgically rotated, described a corollary discharge accompanying every command to move the eyes, and showed that when the discharge and the visual consequence no longer matched the animal circled endlessly, mistaking self-produced motion for world motion (Sperry, 1950).

The two formulations are near-identical in substance and are now used almost interchangeably, efference copy for the copied command and corollary discharge for its downstream perceptual effect. Their shared prediction is testable in humans: a self-produced stimulus should feel attenuated relative to the same stimulus delivered externally, because the former is predicted and cancelled while the latter is not. Blakemore, Wolpert and Frith confirmed exactly this for tactile self-stimulation, showing that a self-administered tactile stimulus is perceived as less intense than an identical externally administered one, and that the attenuation falls off as a delay is introduced between command and touch (Blakemore et al., 1998).

Channels of Sensory Feedback

Feedback is carried by several afferent channels, and no single one is sufficient. Proprioception, signalled chiefly by muscle spindles and Golgi tendon organs, reports muscle length, rate of change and force, and is the dominant source of information about limb configuration (Proske & Gandevia, 2012). Cutaneous mechanoreceptors report contact, slip and the mechanical events of grip, and their signals are used predictively: the hand releases and reloads grip force in anticipation of load changes rather than waiting for the object to slip (Johansson & Flanagan, 2009). Vision supplies a slower but globally accurate estimate of hand and target position; the vestibular system reports head acceleration and orientation; and audition contributes feedback for vocalisation and for contact events. The nervous system does not simply sum these channels. It weights each by its momentary reliability, leaning on vision when proprioception is uncertain and on proprioception when vision is degraded, a reweighting that the Worked Example makes quantitative (Shadmehr et al., 2010).

Forward Models and Optimal Feedback Control

The delay problem forces the nervous system to predict. A forward model is an internal simulation that takes the efference copy and the current state estimate and outputs the sensory consequences the command will produce, before the true feedback arrives (Miall & Wolpert, 1996). Wolpert, Ghahramani and Jordan gave the first direct psychophysical evidence that the human sensorimotor system maintains such a model, showing that estimates of hand position during movement follow the predictions of an internal model that integrates the efference copy with delayed feedback (Wolpert et al., 1995). The forward model also solves the attenuation result: it is the mechanism that predicts and cancels self-produced sensation (Blakemore et al., 1998).

Prediction alone is not control. Optimal feedback control supplies the missing half by specifying how the returning signal should be used: the controller corrects only those deviations that matter for the task and leaves task-irrelevant variability uncorrected, which reproduces the trial-to-trial structure of natural movement (Todorov & Jordan, 2002). Scott set this framework against the physiology, arguing that voluntary motor control is best understood as feedback control operating on internal state estimates rather than as the execution of a stored trajectory (Scott, 2004). Both the model and the controller are learned. Shadmehr and Mussa-Ivaldi showed that reaching in a novel force field is adapted by building an internal model of the limb's altered dynamics, revealed by the after-effects that appear when the field is removed (Shadmehr & Mussa-Ivaldi, 1994), and the cerebellum is central to updating these predictions from sensory prediction error (Bastian, 2006).

Evidence From Deafferentation

The strongest evidence that feedback is necessary comes from its loss. In deafferentation, large-fibre sensory neuropathy strips proprioceptive and tactile afference while leaving the motor pathways and muscle strength intact. Rothwell and colleagues studied one such patient in detail: he could generate accurate isolated movements and normal muscle force, yet could not sustain a constant grip, could not perform tasks requiring sequences of sub-movements without vision, and deteriorated markedly when denied sight of the limb (Rothwell et al., 1982). The dissociation is instructive. What is lost is not the ability to move but the ability to control movement over time, precisely the role a feedback signal plays in the closed loop.

Deafferentation also exposes how heavily control leans on the reweighting described earlier. When proprioception is gone, vision can substitute for many tasks but cannot restore the fast, automatic corrections that spindle feedback normally supplies, because the visual channel is too slow and too attention-demanding (Proske & Gandevia, 2012). The clinical picture matches the computational account: feedback is not a luxury the motor system consults when convenient but a signal it is built around.

Worked Example

When two channels report the same quantity, an optimal estimator combines them in inverse proportion to their variances, so the more reliable channel dominates (Shadmehr et al., 2010). Consider a limb whose position is signalled by vision with a standard deviation of 2 degrees and by proprioception with a standard deviation of 4 degrees. The variances are therefore 4 and 16 square degrees. Each channel's reliability is the inverse of its variance: 1/4 = 0.25 for vision and 1/16 = 0.0625 for proprioception.

The optimal weight on a channel is its reliability divided by the total. Vision receives 0.25 / (0.25 + 0.0625) = 0.80 and proprioception receives 0.0625 / 0.3125 = 0.20. If on one trial vision indicates 10 degrees and proprioception indicates 16 degrees, the fused estimate is 0.80 × 10 + 0.20 × 16 = 11.2 degrees, pulled toward the more reliable visual signal. The variance of the fused estimate is the product of the variances over their sum, 4 × 16 / 20 = 3.2 square degrees, a standard deviation of about 1.79 degrees. That is smaller than either channel alone, which is the central result: integrating feedback channels yields an estimate more precise than any single one. The demonstration below lets these two reliabilities be adjusted and shows the combined distribution narrow as they rise.

Discussion

The reafference principle reframed perception as an active process in which the brain predicts the consequences of its own action and perceives mainly the discrepancy. That idea has proved unusually durable. It explains why a self-produced tickle fails while another person's succeeds, why the visual world stays still across eye movements, and why patients without proprioception can move but cannot control, and it now underlies the predictive-coding view that the brain is fundamentally a prediction machine testing its models against incoming signals. Sensory feedback is the return path of that loop, and its computational description, forward models feeding an optimal feedback controller, has become the standard account of skilled movement (Scott, 2004).

The account also carries a clinical imperative. If feedback is constitutive of control rather than incidental to it, then a prosthetic limb that only receives commands and returns nothing is fundamentally incomplete, however dexterous its actuators. Restoring the afferent half of the loop, closing it around an artificial limb, is therefore not a refinement but the completion of the system (Bensmaia & Miller, 2014). That is the problem the current research front addresses.

Current Directions

The most active application of sensory-feedback theory is the bidirectional prosthesis, a limb that both receives motor commands and returns afferent signals to the nervous system. Two routes are being pursued. Peripheral interfaces stimulate the residual nerves of the stump: Raspopovic and colleagues drove intraneural electrodes in real time from prosthesis sensors and restored tactile feedback that let a user grade grip force and discriminate objects with the device out of sight (Raspopovic et al., 2014). Valle and colleagues later showed that shaping the stimulation to be biomimetic, mimicking the natural firing pattern of tactile afferents, improved the naturalness of the sensation and measurably raised dexterity relative to generic stimulation (Valle et al., 2018). A review of the field frames these gains within the same motor-control theory that governs intact limbs (Sensinger & Dosen, 2020), and a broader survey traces the engineering and clinical state of the art (Raspopovic et al., 2021).

The central route bypasses the periphery entirely. Flesher and colleagues delivered feedback by intracortical microstimulation of human somatosensory cortex, evoking localised, naturalistic sensations referred to the fingers of a paralysed participant, feedback that can be paired with a cortically controlled arm (Flesher et al., 2016). A complementary line targets the sense of movement rather than touch: Marasco and colleagues used tendon vibration to evoke illusory movement perception in prosthesis users and found that supplying this kinaesthetic feedback improved their control of the device (Marasco et al., 2018). Across both routes the open questions are the same, how to encode rich afferent signals in patterned stimulation, how the brain learns to interpret them, and how stable the interfaces remain over years.

Common Misconceptions

Movement is controlled purely by sensory feedback.

Feedback is too delayed to steer fast movement on its own; the nervous system predicts the sensory consequences of action with a forward model driven by an efference copy, and uses feedback to correct the prediction rather than to generate the movement from scratch (Miall & Wolpert, 1996).

The brain treats self-produced and external sensation the same way.

Self-produced stimulation is predicted from the efference copy and attenuated; an identical externally produced stimulus is not. This is why a self-administered touch feels weaker than the same touch delivered by someone else (Blakemore et al., 1998).

Vision can fully replace lost proprioception.

A deafferented person can substitute vision for some tasks but cannot recover the fast, automatic corrections proprioception supplies, because the visual channel is slower and demands attention; control remains impaired even with full sight of the limb (Rothwell et al., 1982).

Glossary

Closed-loop control.

Control in which the returning sensory signal is used to adjust the movement while it is still in progress.

Corollary discharge.

The internal copy of a motor command, in Sperry's terms, whose expected sensory effect is used to interpret the afference that follows.

Deafferentation.

The loss of afferent sensory input, for example through large-fibre neuropathy, leaving motor output intact.

Efference copy.

A duplicate of an outgoing motor command retained centrally so that the sensory consequences of the movement can be predicted.

Exafference.

Sensory stimulation caused by events in the external world rather than by the observer's own movement.

Feedforward control.

Control that specifies a movement in advance from a command and a prediction, without waiting for feedback.

Forward model.

An internal simulation that predicts the sensory outcome of a motor command from an efference copy and the current state.

Kinaesthesia.

The sense of movement and position of the limbs, carried chiefly by proprioceptive afferents.

Optimal feedback control.

A theory in which the controller corrects only task-relevant deviations, leaving irrelevant variability uncorrected.

Proprioception.

The sense of muscle length, velocity and force that reports the configuration of the body.

Reafference.

Sensory stimulation caused by the organism's own movement, distinguished from exafference by the reafference principle.

Sensory prediction error.

The difference between predicted and actual feedback, which drives correction and the updating of internal models.

Sensory reweighting.

Adjusting the relative influence of feedback channels according to their momentary reliability.

State estimation.

The nervous system's ongoing estimate of body and world state, formed by combining prediction with feedback.

Vestibular sense.

The sense of head acceleration and orientation carried by the hair cells of the inner-ear labyrinth.

Key Researchers

Amy J. Bastian (b. 1968). Kennedy Krieger Institute and Johns Hopkins University; she showed how the cerebellum adapts feedforward control from sensory prediction error. ORCID

Sliman J. Bensmaia (1973-2023). University of Chicago; he characterised the neural coding of touch and applied it to somatosensory neuroprosthetics. ORCID

Simon C. Gandevia (b. 1953). Neuroscience Research Australia and the University of New South Wales; he clarified the muscle afferents and central signals underlying proprioception. ORCID

Erich von Holst (1908-1962). Max Planck Institute for Behavioral Physiology; he co-originated the reafference principle relating central commands to peripheral feedback. Wikipedia

Silvestro Micera (b. 1972). EPFL and Scuola Superiore Sant'Anna; he develops neural interfaces and bidirectional prostheses that restore sensory feedback. ORCID

Horst Mittelstaedt (1923-2016). Max Planck Institute for Behavioral Physiology; with von Holst he formulated the reafference principle. Wikipedia

Stephen H. Scott (b. 1964). Queen's University; he developed the optimal-feedback-control account of the neural basis of voluntary movement. Faculty Page

Reza Shadmehr (b. 1963). Johns Hopkins University; he demonstrated internal models of limb dynamics and the role of sensory prediction in motor adaptation. ORCID

Roger W. Sperry (1913-1994). California Institute of Technology; he introduced the concept of corollary discharge and later won the Nobel Prize for split-brain research. Wikipedia

Daniel M. Wolpert (b. 1963). Columbia University; he provided direct evidence for forward models in human sensorimotor integration. ORCID

Frequently Asked Questions

What is sensory feedback? It is the afferent information the nervous system receives as a consequence of its own movement, used to guide and correct action (Miall & Wolpert, 1996).

What is the reafference principle? It is the proposal that the brain distinguishes self-produced stimulation from external stimulation by keeping an efference copy of each command and subtracting its predicted consequence (von Holst & Mittelstaedt, 1950).

How do efference copy and corollary discharge differ? They are near-synonymous formulations from 1950; corollary discharge names the internal signal that accompanies a motor command in Sperry's account of the same mechanism (Sperry, 1950).

Why can a person not tickle themselves? A self-produced touch is predicted from the efference copy and attenuated, so it feels weaker than an identical touch delivered by someone else (Blakemore et al., 1998).

What is a forward model? It is an internal simulation that predicts the sensory consequences of a movement before the actual feedback arrives, letting the system control fast actions despite feedback delays (Wolpert et al., 1995).

What happens when sensory feedback is lost? Deafferented patients retain muscle strength and can move, but lose the ability to control movement over time and deteriorate without vision of the limb (Rothwell et al., 1982).

How does the brain combine different feedback channels? It weights each channel in inverse proportion to its variance, so the more reliable channel contributes more to the fused estimate (Shadmehr et al., 2010).

Can sensory feedback be restored artificially? Yes; intraneural stimulation and intracortical microstimulation can evoke tactile sensations that let prosthesis users grade grip and identify objects (Flesher et al., 2016).

References