Negative Feedback

Negative feedback is a control mechanism in which a system's output is measured, compared against a desired target, and the resulting error signal is used to push the output back toward that target. In machine learning, this principle is most visibly embodied in gradient descent: the loss function measures how far a model's predictions deviate from the ground truth, and the gradient of that loss is used to update weights in the direction that reduces the error. The "negative" in negative feedback refers to the fact that the corrective signal opposes the deviation rather than amplifying it, which is what produces stability rather than runaway divergence.

The mechanism appears throughout modern ML in forms that are not always labeled as feedback. Backpropagation is essentially a structured application of negative feedback across multiple layers of a neural network, propagating error signals backward so that each layer's parameters can be adjusted to reduce the overall loss. In reinforcement learning, the reward signal functions similarly: when an agent's action produces a worse-than-expected outcome, the negative feedback reduces the probability of that action being selected again in similar states, steering policy toward higher-reward behavior over time.

Negative feedback is also central to training stability. Techniques like batch normalization, learning rate scheduling, and adaptive optimizers such as Adam all serve, in part, to regulate the strength and direction of corrective updates so that the feedback loop converges rather than oscillates. Without properly tuned negative feedback, training can become unstable — gradients explode, weights diverge, and the model fails to learn. The challenge in deep learning is not just having a feedback signal but calibrating its magnitude so corrections are neither too aggressive nor too weak.

Understanding negative feedback as a unifying concept helps clarify why so many ML techniques share a common structure: measure error, compute a corrective signal, apply it to reduce future error, and repeat. This loop is the engine of supervised learning, the backbone of reinforcement learning, and a key design consideration in any adaptive system that must improve its behavior over time.