Catch a model taking a shortcut.

A perfect score can hide two explanations

In the original eight examples, shape and background always move together. A rule reading only the background and a rule reading only the shape both score 100%. The data does not distinguish them. A learned system can favor an easy signal if that signal reduces its training loss.

The lab makes the two rules explicit so you can inspect the logic. The shape rule receives the known shape field from our simulated scene; it is not a neural vision model that has solved arbitrary images. The background rule reads only the background field. Their scores demonstrate a problem with evidence, not a contest between trained architectures.

Real examples include an animal classifier using snow instead of anatomy, or a document classifier using a scanner watermark instead of the intended text. More training on the same correlation may reinforce the shortcut.

Change one thing while keeping the answer

Our intervention swaps gold and blue backgrounds. The shape, size, and correct label stay the same. The background rule’s score falls from 100% to 0%; the shape rule keeps its score. Now the two explanations make different predictions, so the test can distinguish them.

This supports a precise conclusion: these two rules respond differently to this background intervention. It does not prove that a real image model is robust to lighting, occlusion, rotation, or unfamiliar shapes. Those claims need additional evidence.

Association is not the effect of an action

An association describes patterns in observations. An intervention deliberately changes something. Seeing that ice-cream sales and sunburn often rise together does not mean buying ice cream causes sunburn; hot sunny weather can influence both. A hidden common cause is called a confounder.

Likewise, observing a background alongside an animal is different from changing the background while holding the animal fixed. Observations alone may not reveal what an action would change. Causal conclusions require suitable assumptions, evidence, and often interventions. This toy scene gives us unusually direct control over which properties stay fixed.

Augmentation has to preserve the intended label

Data augmentation can vary background, brightness, or other irrelevant features so that a model cannot depend on one narrow correlation. But the label-preserving assumption matters. Rotating a 6 can make it look like a 9; changing word order can reverse a sentence’s meaning.

Use varied training examples and separate evaluation examples. Do not choose augmentations solely because they improve the same test set you plan to report. For the module challenge, transfer the idea to sentiment analysis: change the website name without changing the review’s meaning, then compare predictions.