Explain which part of an AI system changes during training and design a small test with unseen examples.
Start with data and a task
Machine learning uses data to fit a model for predictions or generated outputs. The task determines which patterns are useful.
Supervised learning uses examples with target labels or values. A fictional message sorter might learn from messages labeled billing, delivery, or other.
Deep learning uses neural networks with multiple layers. It is one family within machine learning, rather than a synonym for every AI system.
- Training examples
- Fit parameters
- New input
- Model output
Training adjusts parameters; inference uses them on a new input. Retrieval can supply additional context at use time.
Separate learning from use
Training adjusts model parameters using data and a training objective. Those parameters encode learned relationships rather than a searchable filing cabinet of complete sources.
Inference uses a trained model on an input. An ordinary chat response does not necessarily update the model’s parameters.
Retrieval finds external information and supplies it as context. Therefore, adding a document to an answer can change the available evidence without retraining the model.
Understand what a good fit means
Training performance shows how the model handles examples involved in fitting it. However, familiar examples alone cannot establish useful performance on new work.
Generalization means that learned behavior carries to new examples. Evaluation should resemble the intended task while remaining separate from development examples.
Overfitting occurs when a model fits training examples too closely and performs poorly on new inputs. Memorizing yesterday’s messages may fail on tomorrow’s wording.
Work through a fictional sorter
A fictional dataset contains ten billing messages that all include “invoice.” A simple classifier learns that word as its main signal.
A new message says, “I was charged twice,” without using “invoice.” It concerns billing, but the shortcut may miss it.
The lesson is to test varied expressions and ambiguous cases. A perfect result on the original ten messages would not settle this question.
| Change | What changes | What still needs checking |
|---|---|---|
| Add labeled training examples | Learned parameters after training | Performance on new cases |
| Supply a current policy with a request | Context for this inference | Whether the answer uses it faithfully |
| Change the prompt format | Task instructions | Whether required information survives |
Keep system memory separate
Application memory may save notes or retrieve earlier conversations. This does not by itself show that model parameters changed.
Service data policies govern retention and possible later training uses. Check the actual service settings rather than inferring policy from the word “memory.”
A useful explanation names the mechanism: saved record, retrieved passage, current context, or trained parameters. These distinctions make failures easier to investigate.
Practice: sort the mechanism
Your exercise labels three events: fitting from examples, answering a new question, and searching a handbook. Use training, inference, or retrieval.
Your test set adds five fictional billing messages with different wording. Include one that mentions an invoice but asks about delivery.
A strong explanation connects generalization to unseen examples. Explain why a correct answer today does not prove that the model learned permanently.
Compare with the worked answer
Worked answer: Fitting from examples is training; producing an answer is inference; finding handbook passages is retrieval. A delivery question containing “invoice” tests whether the sorter relies on a shortcut. Keep these new cases outside training while using them as an independent evaluation.
What to carry forward
Explain which part of an AI system changes during training and design a small test with unseen examples.
Show your reasoning with the exercise above. Keep the evidence, the decision, and the remaining uncertainty together.