#dataset-bias

Question 1: Single choice

mediumAI-checkedResponsible AI & Fairness

In a biased training set every square appears on the left of the image and every heart on the right; the test set has both shapes everywhere. A supervised shape classifier drops to chance on the test set, while a linear classifier on a SimCLR encoder (pretrained without labels, with random translations among its augmentations) stays accurate. Why?

The supervised model can use horizontal position as a shortcut; translation augmentations make SimCLR features ignore position1
SimCLR was pretrained on the test images, so it has seen every shape at every position2
The supervised model has too few parameters to recognize shapes, so it falls back on guessing3
Linear classifiers cannot overfit, so any linear probe is robust to biases in its training data4

Timed test