Feedback and Error

Imagine a basketball player who practices free throws with their eyes closed and never finds out whether the ball went in. They might feel like they are doing great — but without feedback, they have no way to improve. Feedback is what makes practice meaningful. In machine learning, the feedback signal is called error or loss — a precise mathematical measure of how wrong the model's prediction was. Without it, the training loop you studied in Lesson 2 would have nothing to act on.

What Is Error (Loss)?

Error, also called loss, is a number that measures the gap between what the model predicted and what the correct answer actually was. The larger the error, the worse the prediction. The goal of training is to minimize this number — to drive it as close to zero as possible across all training examples. For a regression problem (predicting a number), a common way to measure error is to simply calculate the difference between the predicted value and the true value. If the model predicted a house price of $300,000 and the real price was $350,000, the error is $50,000. To avoid positive and negative errors canceling each other out, this difference is usually squared — giving a measure called mean squared error. For a classification problem (predicting a category), error is measured differently. The model outputs a probability for each possible category, and the error measures how far those probabilities are from 'perfect confidence in the right answer.' This measure is called cross-entropy loss. The math is more complex, but the idea is the same: small error means the model is confident and correct; large error means it is either wrong or uncertain.

Here is how error works in practice during the training loop from Lesson 2. Example: a model is learning to recognize handwritten digits. It sees an image of a handwritten '7.' Its current (untrained) parameters produce output probabilities like this: 0: 5%, 1: 3%, 2: 8%, 3: 12%, 4: 6%, 5: 9%, 6: 7%, 7: 15%, 8: 18%, 9: 17% The correct answer is 7, but the model only gives it 15% confidence — it actually thinks '8' is most likely. The loss is high. After adjustment, the model might produce: 7: 72%, 8: 10%, 9: 8% (everything else lower) Now 7 wins confidently. The loss is much lower. After thousands more examples and adjustments, the model learns to be highly confident on digits it recognizes clearly. This is feedback turning into learning — exactly the same principle as the basketball player who watches the ball and adjusts their release angle each time.

Error as a Signal for Improvement

The power of error is not just that it measures failure — it is that it gives the model a direction to move. When the loss is high, the gradient descent algorithm (from Lesson 2) uses the loss to calculate how each parameter should change to produce a lower loss next time. Parameters that contributed to the wrong prediction get nudged in the direction of correctness. This feedback loop is what makes machine learning self-correcting. No human needs to look at the model's parameters and decide what to change. The math does it automatically, guided entirely by the loss signal. Over many thousands or millions of training examples, the cumulative effect of these small adjustments is a model that has learned to produce low-loss predictions — in other words, a model that usually gets the right answer.