Masked Autoencoders Quiz

Answers and explanations appear as you take the quiz.

1. What is the core self-supervised task in a Masked Autoencoder?
   • Reconstructing the pixels of patches that were hidden
   • Classifying each image into a labeled category
   • Predicting the next frame in a video sequence
   • Matching an image to its text caption
2. Roughly what fraction of image patches does the original MAE typically mask out?
   • About 15%
   • About 75%
   • About 35%
   • About 95%
3. Why does the MAE encoder process only the visible (unmasked) patches?
   • Because masked patches contain no useful color
   • To prevent the model from ever seeing real images
   • To make training faster and more memory efficient
   • Because the decoder cannot read positional encodings
4. What happens to the decoder after MAE pretraining is complete?
   • It becomes the main classifier
   • It is frozen and used to generate new images
   • It replaces the encoder entirely
   • It is discarded; the encoder is used for downstream tasks
5. Which loss function does MAE typically use for reconstruction?
   • Mean squared error on masked-patch pixels
   • Cross-entropy over class labels
   • Contrastive InfoNCE loss
   • Adversarial discriminator loss
6. The 'asymmetric' design of MAE refers to what?
   • Different image sizes for input and output
   • A heavy encoder paired with a lightweight decoder
   • Using two separate datasets for training
   • Masking only the left half of every image
7. On which large image dataset did the original MAE paper pretrain its main vision models?
   • MS COCO Captions only
   • LibriSpeech
   • ImageNet-1K
   • The Netflix Prize ratings set
8. What does the MAE decoder receive for image patches hidden from the encoder?
   • The original RGB pixels without masking
   • Only class labels from ImageNet
   • A text caption for the image
   • Learned mask tokens with positional embeddings