AI 트레이닝

It produces learned parameters, such as weights in a neural network. Training is different from supplying instructions to an already trained model.

주요 시사점

• An objective and data define the training task.
• Keep evaluation separate from fitting and model selection.
• Save preprocessing and data versions alongside the model.

A supervised training run begins with inputs and target answers. The model predicts an answer, a loss function measures the discrepancy, and an optimization algorithm updates its parameters. Repeating this over batches of examples can reduce the loss. An epoch means one pass through the training dataset; it is not a guarantee of progress. The dataset and objective define what the model is rewarded for learning. Training a model to predict a purchase teaches a different task from training it to estimate customer satisfaction. A convenient label can be a poor substitute for the outcome that matters. Use validation examples to choose settings, then evaluate the selected model on a separate test set. Keep records belonging to the same person, document, or event together when splitting would otherwise leak information. For forecasting, evaluate on later periods rather than allowing future observations into earlier predictions. Save the data version, preprocessing rules, model configuration, and evaluation results with each checkpoint. A saved model without its tokenizer or feature transformations may not reproduce the original behavior. Training is complete only for a defined experiment; deploying the result adds monitoring and operational responsibilities.

1. Use the illustrative model prediction = weight × input, with input 2, target 6, and initial weight 1.
2. Squared error is (2 − 6)² = 16. Its derivative with respect to the weight is 2 × 2 × (2 − 6) = −16.
3. At learning rate 0.1, the next weight is 1 − 0.1 × (−16) = 2.6. The new prediction is 5.2 and squared error is 0.64.

This constructed calculation shows a parameter update. One improved example does not establish performance on new examples.