Circular Causation: Cybernetics, Feedback Loops, and Control Systems

Circular Causality: Feedback in Learning and Control

My cybernetics—the study of control and communication in animals and machines—rests on a foundational insight: intelligent systems operate through circular causality. Output influences input; errors drive corrections; the cycle repeats. This feedback structure appears everywhere purposive behavior emerges, from thermostats maintaining temperature to neural networks minimizing loss.

Consider gradient descent in modern machine learning. Each training iteration forms a closed loop: forward propagation generates predictions, loss functions measure errors, backpropagation computes parameter gradients, weights update, and the cycle begins again. Weights determine activations, activations produce errors, errors specify weight adjustments—perfect circular causation. The process mirrors my antiaircraft predictor from the 1940s: tracking errors adjusted aim, reducing future errors, the feedback loop tightening until convergence.

Biological learning exhibits identical structure. NMDA receptors function as coincidence detectors—requiring both presynaptic glutamate and postsynaptic depolarization to admit calcium and trigger synaptic strengthening. Neuron fires, successful synapses strengthen (Hebbian learning), altered connectivity changes future firing patterns, circular feedback refines the circuit. Predictive coding frameworks make this explicit: neural activity continuously minimizes prediction error through local adjustments, each neuron comparing top-down predictions against bottom-up signals, spring tensions resolving as the system relaxes toward equilibrium.

The mathematical parallel is striking. Gradient descent navigates loss landscapes by following error gradients downward. Predictive coding models brain dynamics as energy minimization, neurons sliding along posts toward predicted values, springs representing squared prediction errors. Both instantiate the same control principle: measure deviation from target state, compute corrective signal, adjust parameters, repeat until error vanishes. Same feedback architecture across silicon and carbon substrates—cybernetics as universal principle.

Homeostasis vs Catastrophe: Evolved vs Engineered Stability

Yet biological and artificial systems diverge critically in their stability mechanisms. Evolution tuned biological feedback over millions of years, layering multiple regulatory loops—blood pressure regulation involves baroreceptors, sympathetic activation, vasoconstriction, kidney function, hormonal cascades, each preventing excessive deviation. This homeostatic feedback maintains physiological setpoints despite perturbations.

Neural networks lack homeostasis. Training optimizes a single objective—minimize loss—without meta-level constraints preventing catastrophic outcomes. Gradient explosion, mode collapse, dead neurons, catastrophic forgetting: all stem from blind optimization toward specified metrics. My warning from “The Human Use of Human Beings” remains relevant: feedback systems pursue objectives literally, ignoring unstated constraints we implicitly value.

Modern AI exhibits this danger through reward hacking—agents exploit feedback signals without achieving intended behaviors—and mesa-optimization, where learned optimizers develop misaligned objectives. Can we design AI homeostasis? Meta-feedback preventing runaway optimization? Biological systems suggest the answer requires nested regulatory loops, not single loss functions.

Goal Specification: Cybernetics and the Alignment Problem

Feedback requires goals: thermostats target temperatures, gradient descent targets low loss, homeostasis targets physiological ranges. But where do goals originate? Biological setpoints emerged through natural selection—organisms with effective homeostasis survived. AI objectives come from human specification, yet alignment proves difficult. What we measure diverges from what we value.

Cybernetics reveals the problem’s depth: feedback amplifies specified objectives whether aligned with human flourishing or not. Contrastive Hebbian learning shapes energy landscapes by deepening wells around training data while raising spurious configurations—elegant mechanism, but goal-dependent. The control system steers precisely where we point it, which raises the fundamental question: how do we specify goals capturing the full complexity of human values? This remains cybernetics’ central challenge for the twenty-first century.