Section 4.7 Machine Learning

Training data for machine learning can come from sources such as sensors, surveys, simulations, or existing data sets. Depending on the task, the data may include labels, which are the correct outputs paired with example data inputs.

Machine learning models use training data to learn patterns that connect inputs to outputs. These patterns come from the examples in the training data and, for supervised learning, the labels attached to them.

Training data is often split into separate sets so the model can learn from one set and then be evaluated on different data it has not seen before. The training set is used to build the model, the validation set is used to adjust model choices and check performance during development, and the test set is used for a final check of how well the model performs on new examples.

In supervised learning, training examples are paired with labels that represent the correct output. Using both examples that match the target and nonexamples that do not match the target helps the model learn boundaries and make more accurate distinctions.

Many machine learning models produce probabilistic outputs, such as a likelihood or confidence score, rather than a guaranteed answer. People decide how to use these scores by choosing a threshold, which determines when the system accepts a prediction automatically and when it requires more checking or human review.

When using a threshold, a system can make two kinds of mistakes. A false positive occurs when the system incorrectly accepts something that should be rejected. A false negative occurs when the system incorrectly rejects something that should be accepted. Changing the threshold usually reduces one kind of error while increasing the other, so developers choose thresholds by considering which error is more harmful for the user and the situation. Developers may also add human review or additional checks to reduce harm.

Training sets that include diverse examples and counterexamples help a model generalize, meaning it performs well on new inputs and varied real-world situations, not just the examples it learned from.

A model can appear to work well during development but still perform poorly on new inputs if it learns patterns that do not generalize beyond the training examples. Testing a model on new data helps evaluate whether its patterns transfer beyond the training sets.

Training data may be processed before a model is trained to decide which examples are included and how they are organized. Sampling can reduce data set size or balance the number of examples across categories. Sampling decisions can improve performance for underrepresented cases, but they can also reduce accuracy if important cases are removed or under-sampled.

Training data can expose sensitive information. Privacy protection includes minimizing, limiting access to, and protecting sensitive data during storage and sharing.

Bias can occur when training data does not represent the people or situations the model will be used for, or when labels reflect human assumptions. Collecting more data does not necessarily reduce bias if the same sampling methods or labeling choices continue.

Training data should reflect the intended users and real-world contexts where the model will be used. If the data set differs from actual use, the model can perform poorly or unevenly across users even if it appears accurate during testing. Evaluating training data includes checking whether it represents the people, environments, languages, and situations the system is meant to support, and testing with representative user scenarios to reduce confusion and unequal performance.