Virtual Labs

Autoencoders for Representation Learning

1. After training, what does the encoder of an autoencoder produce?

a: A compressed latent representation of the input Explanation

Explanation

b: The final classification label Explanation

Explanation

c: The reconstruction loss value Explanation

Explanation

d: Random noise Explanation

Explanation

2. In the denoising autoencoder experiment, what should the target output be?

a: The noisy input image Explanation

Explanation

b: A completely different image Explanation

Explanation

c: The original clean image Explanation

Explanation

d: The latent representation Explanation

Explanation

3. What type of noise is added to images when training the denoising autoencoder?

a: Salt-and-pepper noise only Explanation

Explanation

b: Gaussian noise Explanation

Explanation

c: No noise is added Explanation

Explanation

d: Colour channel noise Explanation

Explanation

4. If the reconstruction loss on the validation set decreases during training, it indicates:

a: The model is learning to better reconstruct unseen inputs Explanation

Explanation

b: The model is overfitting severely Explanation

Explanation

c: The latent dimension is too large Explanation

Explanation

d: The dataset is corrupted Explanation

Explanation

5. What activation function is used in the final decoder layer?

a: ReLU Explanation

Explanation

b: Tanh Explanation

Explanation

c: Sigmoid Explanation

Explanation

d: Softmax Explanation

Explanation

6. When visualising the latent space, what does clustering of similar items indicate?

a: The model failed to learn Explanation

Explanation

b: The autoencoder has learned meaningful representations that group similar items together Explanation

Explanation

c: The training data was insufficient Explanation

Explanation

d: The latent dimension is too large Explanation

Explanation

7. What does the bottleneck layer force the autoencoder to learn?

a: How to add more noise to images Explanation

Explanation

b: Only the pixel colours of images Explanation

Explanation

c: A compressed representation capturing essential features of the input Explanation

Explanation

d: The exact copy of input without any compression Explanation

Explanation

8. Why is batch normalisation used in the autoencoder architecture?

a: To reduce the number of training epochs to zero Explanation

Explanation

b: To stabilise training by normalising layer inputs and reducing internal covariate shift Explanation

Explanation

c: To eliminate the need for activation functions Explanation

Explanation

d: To automatically add noise to images Explanation

Explanation

9. What does PSNR (Peak Signal-to-Noise Ratio) measure?

a: Training speed Explanation

Explanation

b: Reconstruction quality in decibels, with higher values indicating better quality Explanation

Explanation

c: Number of model parameters Explanation

Explanation

d: Latent space dimensionality Explanation

Explanation

10. If the autoencoder reconstructs training images perfectly but performs poorly on test images, what problem is occurring?

a: Underfitting Explanation

Explanation

b: The latent dimension is too small Explanation

Explanation

c: The learning rate is too low Explanation

Explanation

d: Overfitting Explanation

Explanation