Test-Time Augmentation Quiz

Answers and explanations appear as you take the quiz.

1. At inference time, what does test-time augmentation do with a single input?
   • Running the fixed model on several transformed copies of an input and combining the outputs
   • Retraining the model on augmented data each time it is queried
   • Adding noise to the model's weights before each prediction
   • Compressing the model so it runs faster at inference
2. When can pooling predictions from transformed views reduce prediction variance?
   • When every view changes the true class
   • When the transforms preserve the label and the view errors are at least partly uncorrelated
   • When all predictions are identical by design
   • When the model is retrained on each transformed copy
3. What is the main cost of using TTA?
   • It requires labeled test data
   • It permanently alters the trained weights
   • It increases inference compute because the model runs multiple times per sample
   • It reduces model accuracy on clean inputs
4. Which augmentation choice could actually HURT a digit classifier?
   • Small brightness changes
   • Tiny random crops
   • Adding a little Gaussian noise
   • Horizontally flipping a '6' so it can look like a '9'
5. In segmentation, what extra step is required before averaging TTA outputs?
   • Inverting each geometric transform so the predicted masks realign with the original pixels
   • Retraining the decoder for each view
   • Discarding all flipped views
   • Converting masks to bounding boxes
6. For a classification model, what may be aggregated across test-time views?
   • The model’s training labels
   • The predicted probabilities or logits from the views
   • The input image file names
   • The optimizer momentum buffers
7. Does standard test-time augmentation require retraining the model on new labels?
   • Yes; each view needs a newly trained model
   • Yes; it requires extra class labels for every view
   • No; it reuses fixed weights and evaluates transformed inputs at inference
   • No; it replaces the model with a lookup table
8. A segmentation model predicts masks on flipped and cropped views. What must happen before pixelwise pooling?
   • Average masks before undoing the transforms
   • Discard the original image coordinates
   • Apply a different label to each pixel view
   • Invert the geometric transforms so all masks align with the original image coordinates