Is error-driven learning always supervised?

It commonly appears in supervised learning, but prediction errors also occur in other learning arrangements.

Does a chatbot learn permanently when I correct one answer?

Not necessarily. A conversation can affect current context without changing the model's trained parameters.

Is backpropagation the same as the whole learning process?

No. It computes gradients in suitable models; data, objectives, optimization and evaluation are additional parts.

AI Glossary / AI and machine learning / Intermediate

Error-driven learning

Definition

Error-driven learning adjusts a model using a discrepancy between its output and a desired or observed outcome. The error supplies information about how the model's future responses should change under the chosen learning rule.

What it means

A learner needs some basis for improvement. In an error-driven process, it makes a prediction, compares that prediction with a target and uses the difference to guide an update.

The error may be a simple numerical difference, such as predicting eight units when the target is ten. More commonly, a loss function translates the mismatch into a quantity suitable for the training objective. Different tasks require different definitions of error.

The phrase is broad. It can describe learning rules in machine learning and appears in other fields with their own meanings. It should not be treated as the name of one universal algorithm or as proof that a system understands why it was wrong.

For business users, the essential question is where the corrective signal comes from. If targets are incorrect, delayed or biased, learning from them can reinforce the wrong pattern rather than improve useful performance.

How it works

The model receives an input and produces an output. A target or feedback signal is obtained, and the training objective measures a discrepancy between the two.

A learning rule determines how that discrepancy affects the model. In a simple linear learner, the update can depend on the input, prediction error and learning rate. In differentiable multilayer networks, backpropagation calculates derivatives that an optimizer can use.

The update changes parameters so that future outputs may better match the training objective. Repeating this process across examples can reduce loss, but noisy data and interacting updates mean every individual step need not improve every case.

Performance on held-out data is monitored separately. Memorizing corrections to training examples is not the same as learning a pattern that works on new requests, products or conditions.

Three practical examples

Demand estimation: a model uses differences between predicted and observed demand during a defined training process.
Image classification: known training labels provide the target against which a model's predicted class distribution is evaluated.
Quality prediction: reviewed examples of acceptable and defective items provide feedback for updating a classifier, subject to label-quality checks.

Think of it this way

Think of practicing a serve in tennis and using information about where the ball landed to adjust the next attempt. Feedback about the result guides a change rather than merely rewarding repetition.

The analogy has a limit. A person can interpret causes and question misleading feedback. A mathematical update follows the specified signal, so an incorrect target can push a model in an unhelpful direction.

Worked example

Calculating a simple correction

Suppose a teaching model predicts y as w multiplied by x. Let x equal two, w equal three and the target equal eight. The prediction is six, so target minus prediction equals two.

Using the illustrative update w_new equals w plus learning rate multiplied by error multiplied by x, and a learning rate of 0.1, the new weight is 3 + 0.1 × 2 × 2, or 3.4.

For the same input, the new prediction is 6.8, closer to eight in this example. This update corresponds to a simple squared-error setting with a suitable convention; it is not the formula for every learning algorithm.

If the target of eight were a recording error, the apparently successful correction would move the model toward bad information. The arithmetic illustrates the learning mechanism, not the trustworthiness of the target.

Illustrative case study

A fictional service company retrains a model using staff corrections to its routing suggestions. A review finds that some corrections reflect temporary staffing shortages rather than the true category of the request.

The team separates request category from the queue that happened to receive the work. It clarifies the training target and reviews a sample of labels before continuing updates.

The invented lesson is that an observed correction can contain several meanings. Error-driven learning needs a signal aligned with the intended task, not an unexamined record of whatever action someone took.

Common misconceptions and mistakes

Any correction is a reliable learning target. Feedback may be mistaken, strategic or unrelated to the intended prediction.
A lower training error proves better real-world performance. Independent evaluation is needed to detect overfitting and changed conditions.
The model understands its mistake. The update uses numerical or symbolic feedback according to a learning procedure; understanding is not established by the correction.

Business use

The concept is useful whenever a supplier promises that a system will improve from feedback. Ask precisely which feedback is collected, how it becomes a target and when it changes the deployed model.

Separate ordinary user interactions from approved training data. Review feedback quality, version updates and evaluate changes before assuming that more corrections inevitably produce a better service.

Check your understanding

Staff route a request to a different queue because one team is absent. Why might that action be a poor target for learning the request's true category?

Answer: It reflects a temporary operational constraint rather than the category itself. Training on it without distinction can teach the model the wrong relationship.

Related terms

Learn next

Read loss and learning rate. Trace one proposed feedback signal back to its origin and ask whether it represents the outcome the model should learn.

Sources

Rumelhart, Hinton and Williams, Learning representations by back-propagating errors, 1986; author publication record: https://www.cs.toronto.edu/~hinton/backprop.html. Record checked 12 October 2026. The simple linear update and business scenarios are original teaching examples.

Previous: Error cost function Next: Error exponent