Machine learning derives rules from examples instead of having a person write them. In ordinary programming you supply the logic and the data, and the computer returns answers. In machine learning you supply the data and the answers, and the computer works out the logic. That logic is a model, a set of numbers fitted to patterns in past examples.

The difference earns its cost when the rules are too messy to write down. Nobody can enumerate every way a photo can show a cat. But you can collect labeled photos and let training find the pattern for you.

What you give up is certainty. A hand-written rule fails in ways you can read in the source code. A model fails in ways you have to measure. You also sign up for data collection, labeling, and retraining as the world shifts.

Supervised learning learns from examples that carry a correct answer, called a label. You hand it past emails marked spam or not spam, and it learns the mapping. Unsupervised learning gets rows with no answers attached at all. It looks for structure in the data itself, like groups of similar customers.

The practical split is usually about what you already have. Labels cost money, because someone has to read every email and mark it. When nothing is labeled, unsupervised methods still tell you something useful about the shape of your data.

Checking the result differs too. A supervised model has a right answer to score against, so error is a number. Unsupervised output needs a person to look at the clusters and decide whether they mean anything.

The training set is what the model actually fits. Its rows adjust the parameters until error on those rows comes down. The validation set is where you make choices: which model, which settings, how long to train. You score candidates on it and keep the winner.

The test set gets used once, at the end. It estimates how the chosen model will do on data nobody tuned against. That is a different question from validation: not which option is best, but how good the winner really is.

A common split is 60 percent training, 20 validation, 20 test. Large datasets can spare far less for the last two. Skip the split and every number you report describes rows the model already memorized.

Overfitting means the model learned the training rows too well, noise included. It reproduces quirks that will not repeat, so new data goes worse than old. Underfitting is the opposite failure. The model never captured the real pattern, so it does poorly everywhere, training rows included.

Overfitting comes from more flexibility than the data can support. Typical causes are too many parameters, too few rows, or too many passes over the same data. Features that quietly identify individual rows do it too.

Underfitting comes from the reverse. The model is too simple for the pattern, the features leave out what matters, or training stopped early.

Naming them separately matters because the fixes point in opposite directions. Adding capacity to an overfit model makes it worse, and simplifying an underfit one does the same.

Bias is error that comes from a model's assumptions being too simple for reality. A straight line fitted to a curved relationship is biased. It misses the shape no matter how many rows you feed it.

Variance is how much the fitted model changes when you swap in a different training sample. A high-variance model chases the particular rows it saw. Two samples drawn from the same source give you two noticeably different models.

The dartboard picture is the fastest way in. Bias is aiming at the wrong spot, and variance is a shaky hand. You can be consistently wrong, randomly wrong, both, or neither.

Keeping the two words separate matters because they point at different repairs. Bad aim and a shaky hand are not fixed by the same change.