Autoencoders Quiz

Answers and explanations appear as you take the quiz.

1. What is the purpose of the 'bottleneck' (latent code) in an autoencoder?
   • To force the network into a compact representation that captures essential structure
   • To add more parameters and increase model capacity
   • To store labels for supervised classification
   • To slow down training for better stability
2. Why are autoencoders considered self-supervised?
   • They require huge labeled datasets
   • The training target is the input itself, so no external labels are needed
   • A human labels each reconstruction by hand
   • They only work on pre-classified images
3. How does a denoising autoencoder differ from a standard one?
   • It removes the decoder entirely
   • It only works with text data
   • It is trained to reconstruct a clean input from a corrupted version
   • It requires no bottleneck layer
4. What makes a variational autoencoder (VAE) able to generate new data?
   • It memorizes every training example
   • It uses no decoder
   • It replaces neural networks with decision trees
   • It learns a smooth, probabilistic latent space you can sample from
5. A linear autoencoder trained with mean-squared error closely resembles which classic technique?
   • Principal component analysis (PCA)
   • K-means clustering
   • Decision trees
   • Logistic regression
6. Why are autoencoders well suited to anomaly detection?
   • They label anomalies during training
   • They reconstruct common patterns well but show large errors on rare, unseen patterns
   • They ignore reconstruction error entirely
   • They only output binary classifications
7. Which reconstruction loss does the guide associate with binary-pixel data?
   • Triplet margin loss
   • Temporal-difference error
   • Cross-entropy
   • Cosine annealing loss
8. Why can nonlinear activations make an autoencoder more expressive than a linear one?
   • They force the latent code to equal the input size
   • They remove the decoder from the model
   • They guarantee zero reconstruction error on unseen data
   • They allow it to learn nonlinear or curved structures in the data