Is a learning rule always supervised?

No. Rules can use unlabeled structure or reward signals instead of labeled targets.

Can the same model use different learning rules?

Yes, provided the rules are compatible with its representation and training objective.

Must every update improve accuracy?

No. Rules often optimize a surrogate objective, stochastic updates can fluctuate, and improvements on training data may not transfer to evaluation data.

AI Glossary / Deep learning / Beginner

Learning rule

Definition

A learning rule is a procedure that specifies how a model changes its adjustable parameters or internal state in response to data or feedback. In a neural network, it often determines how connection weights are updated. Learning rules can use prediction errors, correlations between activities, rewards, or other signals. They describe the mechanism of adaptation, not a guarantee that every update improves real-world performance.

What it means

A model architecture says what computations are possible with its current parameters. A learning rule says how those parameters change. The loss function defines an objective, while a learning rate controls the size of some kinds of update. These concepts work together but are not interchangeable. A rule may be local, using information available around one connection, or depend on information propagated through a larger network. Different rules can train the same architecture, sometimes producing substantially different results.

How it works

A familiar example is gradient descent: calculate the derivative of a chosen loss with respect to each parameter, then subtract the learning rate times that derivative. Backpropagation efficiently computes the derivatives in many neural networks; the optimizer uses them to change parameters. Other rules adjust connections according to co-activity or move a winning prototype toward an observation. To specify a rule fully, state its feedback signal, update equation, learning-rate choices, timing, and any constraints or stored optimizer state. A label such as “adaptive learning” is not enough.

Three practical examples

Demand prediction: update model weights using differences between predicted and observed quantities under a specified regression loss.
Unsupervised mapping: move selected prototype vectors toward incoming observations according to a competitive learning rule.
Interactive optimization: use measured rewards to adjust a policy through a defined reinforcement-learning update, while distinguishing the reward from the update mechanism.

Think of it this way

Think of a cook adjusting a recipe after tasting a batch: the rule specifies which ingredient changes and by how much. The analogy captures feedback-driven adjustment. Its limit is that a mathematical rule follows a defined signal and cannot independently know whether that signal represents the broader goal, and a change that helps one batch may hurt another.

Worked example

Consider a one-parameter predictor y-hat = w times x, with one training example x = 2 and target y = 6. Let w initially equal 1 and use loss L = one-half times (y-hat minus y) squared. The prediction is 2, the error is -4, and the loss is 8. The derivative with respect to w is (w times x minus y) times x, giving -8.
With learning rate 0.1, gradient descent gives new w = 1 minus 0.1 times (-8) = 1.8. The new prediction is 3.6, the error is -2.4, and the loss is 2.88. The same gradient with learning rate 1 instead produces w = 9, prediction 18, and loss 72. The update direction began as a descent direction, but an excessively large step overshot the favorable region. This example demonstrates both the rule and the effect of its step size. Improvement on this one example says nothing by itself about unseen inputs or a larger training set.

Illustrative case study

A fictional service-planning team notices that a forecasting model reacts sharply to each day's newest record. It examines the update rule, the scale of its inputs, and the learning rate rather than assuming the architecture is faulty. The team compares smaller steps and batched updates on historical training periods, then evaluates on later periods reserved for testing. It keeps a simple baseline and records when updates occur so it can distinguish an adaptation problem from a change in the underlying demand pattern.

Common misconceptions and mistakes

Backpropagation is not, by itself, a complete parameter-update policy: gradients still need an optimizer and configuration. A learning rate is a setting within many rules, not the entire rule. A lower training loss does not guarantee better generalization, and a noisy minibatch update may temporarily raise loss elsewhere. Unsupervised learning still has a defined adaptation mechanism. Biological inspiration also does not prove that an artificial update rule faithfully models a brain or is suitable for every business problem.

Business use

Understanding the learning rule helps teams plan retraining, compare reproducibility, and diagnose unstable adaptation. Ask which parameters are allowed to change, which data or rewards drive the changes, and whether updates happen offline or during operation. Preserve optimizer state when it is needed to resume training consistently. Evaluate the consequences of learning-rate schedules, data order, and stopping criteria using an appropriate validation design. For a deployed system, make model versions and update history traceable so a change in behavior can be investigated.

Check your understanding

In the worked example, why does learning rate 1 increase the loss even though the gradient points toward lower loss locally?

Answer: The local descent direction does not guarantee improvement for an arbitrarily large finite step. The update jumps past the useful region, producing weight 9 and loss 72 instead of the initial loss 8.

Related terms

Learn next

Start with one-parameter gradient descent and calculate an update by hand. Then separate the objective, gradient calculation, and optimizer in a small training loop. Compare a basic gradient update with a prototype update to see why “learning” includes several mechanisms. Finally study validation and stopping rules, which determine whether adaptation serves the intended task.

Sources

Goodfellow, Bengio, and Courville, Machine Learning Basics, including stochastic gradient descent: https://www.deeplearningbook.org/contents/ml.html Goodfellow, Bengio, and Courville, Optimization for Training Deep Models: https://www.deeplearningbook.org/contents/optimization.html Kohonen and Honkela, Kohonen networks: https://www.scholarpedia.org/article/Kohonen_networks

Previous: Learning rate Next: Learning task