Hoe AI leert

The aim is to perform well on new examples, not simply to remember the training examples; some AI systems use explicit rules and do not learn this way at all.

Key takeaways

• Training changes the model; inference uses it.
• Keep evaluation examples separate from the examples used to choose or train the model.
• Choose metrics that reflect the cost of mistakes, not only a large accuracy number.

In supervised learning, training examples pair inputs with target outputs. The model makes a prediction, a loss function measures how far that prediction is from the target, and a training algorithm changes the model to reduce the loss. Neural networks commonly use gradient-based optimization, but not every learning algorithm uses gradients. Validation data helps developers choose settings and compare candidate models. A held-out test set provides a separate estimate of performance after those choices are made. Repeatedly choosing models based on the test set weakens that separation. If the same person, document, or near-duplicate example appears on both sides of a split, the result can look better than performance on genuinely new data. Other learning setups use different signals. Unsupervised learning looks for structure without a target label for every example. Self-supervised training creates prediction tasks from the data itself, such as predicting text that follows a context. Reinforcement learning uses feedback about actions and outcomes. In every case, the training objective is a useful proxy, not a complete definition of what people want. After training, inference is the use of the model to produce an output. Supplying an example in a prompt can change the current response without updating the model's learned weights. Whether a service later uses a conversation for training is a separate product and data-policy question.

1. Imagine 100 test messages: 10 are spam and 90 are legitimate. A system that never flags spam is 90% accurate but catches none of the spam.
2. Another system flags 20 messages. Eight really are spam and 12 are legitimate. It misses two spam messages.
3. Its accuracy is 86%, precision is 8/20 = 40%, and recall is 8/10 = 80%. Decide whether catching eight spam messages is worth wrongly flagging 12 legitimate messages.

These are invented counts for an arithmetic example, not a benchmark result. They show why a single metric cannot determine whether a model is fit for a task.