Concept
Feedback
Routing a system's output back to its input so it corrects itself — long used in steam-engine governors, brought to amplifiers as negative feedback in 1927, and since found running almost everywhere in the nervous system.
From the EE side
From the neuro side
In 1927 Harold Black of Bell Telephone Laboratories needed repeaters that could be chained by the dozen without their errors adding up. Black’s insight was that you could throw away most of an amplifier’s gain and buy accuracy with it. Biology appears to have reached the same bargain repeatedly, and usually at several nested timescales at once.
The amplifier’s bargain
Subtract a fraction of an amplifier’s output from its input, and its gain becomes
With the loop gain large, approaches , set by passive parts, and drift and distortion in the tubes are divided by . Black’s 1934 paper, “Stabilized Feedback Amplifiers”, reports gain moving less than 0.01 dB as the plate supply swung from 240 to 260 V, against about 0.7 dB for a conventional design.
The price is stability. Enough phase shift turns subtraction into addition, and if the loop gain is still above one at that frequency the amplifier can sing, which is how oscillators are built on purpose. Harry Nyquist’s “Regeneration Theory” of 1932 — regeneration being his word for feedback — gave the test: plot the loop’s frequency response and see whether it encloses a critical point.
The nervous system’s version
In the Hodgkin–Huxley model, depolarisation raises sodium conductance and the sodium current depolarises further — positive feedback, the upstroke. Sodium inactivation and a delayed rise in potassium conductance pull the membrane back: negative feedback, late. The voltage clamp that measured those conductances is Black’s amplifier pointed at an axon, holding the voltage still so the sodium loop cannot run. The patch clamp is the same loop at picoamps.
Slower loops sit on top: purely spinal circuits correct a stretched arm muscle 20–45 ms later. In 1998 Gina Turrigiano and colleagues pushed cultured cortical neurons to fire harder; over 48 hours their synapses scaled down and firing returned close to baseline. That slow gain control was proposed as a brake on Hebbian learning, though later modelling suggests it is too slow to stop a runaway on its own. Dopamine neurons appear to signal the error itself: reward prediction error.
Where the readings meet, and where they part
Arturo Rosenblueth, Norbert Wiener and Julian Bigelow made the crossing explicit in 1943: purposeful behaviour could be treated as requiring negative feedback, and a cerebellar patient’s tremor, growing as the hand nears its target, looked to them like undamped feedback. Wiener’s Cybernetics named the field in 1948. In 1958 Lawrence Stark and Tom Cornsweet took the pupil reflex’s frequency response, which Stark and Philip Sherman had measured in 1957, and put it on a Nyquist plot: loop gain 0.12 where the phase lag reached 180°, so raising the gain past one should make the pupil oscillate at about 72 cycles a minute. A spot of light on the pupil’s margin, a clinical trick from 1944, does exactly that, near the predicted rate.
The analogy breaks at the reference. In Black’s amplifier the reference is a terminal you can probe. In a nervous system the set point is often implicit: a resting potential is a balance of permeabilities, not a command, and Andrej Romanovsky argued in 2007 that body temperature is held by independent effector loops with no single set point. Closed-loop deep brain stimulation, like responsive neurostimulation, has to supply what the biology never declared — a signal such as beta-band power, a rule such as a hand-set threshold, and a stimulator it can command.
Nearby concepts
All topics under Feedback
- Responsive neurostimulation 2013 A skull-mounted implant that reads the cortex at a seizure focus and stimulates it when a detector fires — closed-loop brain stimulation, approved in the US in 2013.
- Reward prediction error 1997 Dopamine neurons signal not reward but the difference between reward received and reward expected — an error signal, of exactly the kind a feedback controller runs on.
- Neuromorphic computing 1990 Building computation from analog devices in their physics-limited regimes and from spikes rather than clocked arithmetic, on the argument that the brain's efficiency comes from its style of computing.
- Deep brain stimulation 1987 Chronically implanted electrodes delivering continuous high-frequency pulses, which treat Parkinsonian symptoms effectively while nobody fully agrees on why.
- The patch clamp 1976 A glass pipette sealed against a membrane made the current through single ion channels measurable, turning channels from inference into instrument readings.
- Lateral inhibition 1957 Neighbouring receptors inhibit one another, so a sheet of them exaggerates edges and suppresses uniform regions — measured in the horseshoe crab's eye, later rebuilt in silicon.
- The Reichardt motion detector 1956 Delay one receptor's signal, multiply it by its neighbour's and subtract the mirror image: a motion detector inferred from a beetle's turning in 1956, since built in silicon and found in the fly.
- The Hodgkin–Huxley model 1952 Circuit theory borrowed to explain the axon as a capacitor with voltage-dependent conductances — then handed back decades later as the template for analog silicon neurons.
- Hebbian learning 1949 Neurons that fire together, wire together — a synapse strengthens when the cells it joins are repeatedly active together, using nothing but signals already present at the synapse.
- The voltage clamp 1949 A feedback amplifier that holds membrane potential at a commanded value and reports the current needed to do it, which is what turned excitability into a measurable quantity.
- Central pattern generators 1911 Neural circuits that produce the rhythms of walking, swimming, flying and breathing from an input with no rhythm in it, leaving sensory feedback to adjust the pattern rather than create it.
- Programmable stimulation waveforms — Synthesising the stimulus pulse with a modular converter instead of inheriting whatever the output stage rings at, which turns shape and timing into experimental variables.