Computing / Lesson 05

Learning systems and evaluation

Treat a trained model as one component in an evidence-producing system.

Reading time
64 minutes
Evidence
unseen
Release
0.1-preview
Review
Subject review pending

orient

Why this idea had to exist

Machine learning replaces hand-written rules with behavior fitted from examples. That flexibility makes evaluation central: a model can fit its training distribution while failing where deployment conditions differ.

intuition

Build a picture you can reason with

Training is rehearsal, validation is direction, and testing is opening night. Reusing the audience's reactions to rewrite the play turns the test into more rehearsal. Separate data roles preserve the meaning of the final evidence.

formalize

Give the intuition a precise edge

A learning algorithm selects parameters using training data. Hyperparameters and decisions use validation evidence. A held-out test estimates performance for a specified sampling process. Subgroup, shift, calibration, and error-cost analyses test assumptions hidden by one average metric.

work through

Follow the decisions, not just the symbols

For a rare failure class, 99 percent accuracy can come from always predicting the common class. Build a confusion matrix, inspect recall and precision, and choose a threshold based on the real cost of misses and false alarms.

experiment

Change one thing and watch the model answer

Move a classification threshold and watch false positives and false negatives trade places. Then change class prevalence and observe why precision changes even when sensitivity and specificity stay fixed.

retrieve

Close the page and reconstruct it

Answer before opening the explanation. Retrieval is evidence only when the answer is produced without a hint.

Why should the test set not guide model selection?

Using it to choose the model adapts decisions to that sample, so its score no longer estimates performance on untouched data.

transfer

Move the idea into a new setting

Write an evaluation plan for a model used across common, code-heavy, and dialogue-heavy environments, including one worst-environment measure.

reflect

Leave with a diagnostic habit

A score answers only the question encoded by its data, metric, and decision threshold. Write that question beside every result.

Source record

Follow the idea back.

Reference
Machine Learning Crash Course
Publisher
Google for Developers
License
CC BY 4.0
Accessed
2026-08-12
URL
https://developers.google.com/machine-learning/crash-course