Autoencoders quiz

Question 1

What is an autoencoder in machine learning?

Accepted Answer

A neural network trained to reconstruct its own input through a compressed code

Answer

A network that classifies images into fixed categories

Answer

An algorithm that always requires labeled data to learn

Answer

A method for sorting numbers in ascending order

Question 2

What are the two main components of an autoencoder?

Accepted Answer

Encoder and decoder

Answer

Generator and discriminator

Answer

Actor and critic

Answer

Tokenizer and embedder

Question 3

What is the 'bottleneck' or latent space in an autoencoder?

Accepted Answer

The compressed middle layer holding the latent code

Answer

The output layer that produces class probabilities

Answer

The loss function used during training

Answer

A regularization penalty added to the weights

Question 4

Why are autoencoders considered unsupervised (or self-supervised) learning?

Accepted Answer

The input itself serves as the target, so no labels are needed

Answer

They use a teacher network to provide labels

Answer

They require a fully labeled dataset

Answer

They only work on classification labels

Question 5

What loss function is commonly used to train an autoencoder on continuous (real-valued) data?

Accepted Answer

Mean Squared Error (reconstruction loss)

Answer

Hinge loss

Answer

Categorical accuracy

Answer

Policy gradient loss

Question 6

What is the key difference between an undercomplete and an overcomplete autoencoder?

Accepted Answer

Undercomplete has a smaller-than-input code; overcomplete has an equal/larger code

Answer

Undercomplete uses labels; overcomplete does not

Answer

Undercomplete has no decoder; overcomplete has two decoders

Answer

They are identical except in name

Question 7

Why might an autoencoder learn just to copy its input, and how is this prevented?

Accepted Answer

By using a bottleneck, regularization, or noise to prevent trivial copying

Answer

By making the network deeper, which guarantees learning

Answer

By using only one neuron in every layer

Answer

By removing the decoder entirely

Question 8

What is a denoising autoencoder?

Accepted Answer

One trained to reconstruct clean input from a noisy/corrupted version

Answer

One that adds noise to its output for variety

Answer

One that only works on audio signals

Answer

One with no encoder, only a decoder

Question 9

How are autoencoders used for dimensionality reduction?

Accepted Answer

The low-dimensional latent code is used as a compressed feature representation

Answer

The decoder output is used as the reduced data

Answer

They increase dimensions to add detail

Answer

They cannot reduce dimensionality at all

Question 10

How does an autoencoder differ from PCA (Principal Component Analysis)?

Accepted Answer

Autoencoders can learn nonlinear compressions while PCA is strictly linear

Answer

PCA is always more accurate than any autoencoder

Answer

Autoencoders require labels but PCA does not

Answer

They are mathematically identical in all cases

Question 11

What is reconstruction error and why is it important?

Accepted Answer

The difference between input and its reconstruction; the value being minimized

Answer

The number of layers in the encoder

Answer

The time taken to train the model

Answer

The size of the latent vector

Question 12

How can autoencoders be used for anomaly detection?

Accepted Answer

Anomalies produce high reconstruction error and exceed a chosen threshold

Answer

Anomalies produce zero reconstruction error

Answer

The decoder labels anomalies directly

Answer

Anomalies are detected by the encoder's bias terms

Question 13

What is a variational autoencoder (VAE) at a high level?

Accepted Answer

A generative model that encodes inputs as distributions and can sample new data

Answer

A classifier that outputs class labels only

Answer

An autoencoder with no decoder

Answer

A model that only removes noise from images

Question 14

In a variational autoencoder, what role does the KL-divergence term play in the loss?

Accepted Answer

It keeps the latent distribution close to a standard normal, regularizing the space

Answer

It maximizes reconstruction error

Answer

It counts the number of training samples

Answer

It replaces the decoder network

Question 15

What is the 'reparameterization trick' used in VAEs?

Accepted Answer

Expressing a sample as mean + std*epsilon so gradients can flow during backprop

Answer

Reusing the same parameters for encoder and decoder

Answer

Reducing the number of parameters by sharing weights

Answer

Randomly resetting weights each epoch

Question 16

What is a sparse autoencoder?

Accepted Answer

One with a penalty encouraging most latent neurons to be inactive per input

Answer

One that uses only sparse matrices for speed

Answer

One trained on very small datasets

Answer

One that has no hidden layers

Question 17

What does the encoder function compute in an autoencoder?

Accepted Answer

It maps the input x to a compressed latent code z

Answer

It maps the latent code back to the input

Answer

It computes the loss directly

Answer

It generates random noise

Question 18

What does the decoder function compute in an autoencoder?

Accepted Answer

It maps the latent code z back to a reconstruction of the input

Answer

It compresses the input into the code

Answer

It calculates the gradient of the weights

Answer

It selects the learning rate

Question 19

Why is a nonlinear activation function important in deep autoencoders?

Accepted Answer

They let the network learn nonlinear mappings beyond what PCA can capture

Answer

They make training faster but add no expressive power

Answer

They are only needed in the loss function

Answer

They replace the need for a decoder

Question 20

What is a convolutional autoencoder and when is it used?

Accepted Answer

An autoencoder using convolutional layers, suited to image data

Answer

An autoencoder that only works on text

Answer

An autoencoder without any hidden layers

Answer

A model that uses convolution only in the loss

Question 21

Can a basic (vanilla) autoencoder reliably generate brand-new, realistic samples? Why or why not?

Accepted Answer

No, its latent space is not smooth, so random points decode poorly

Answer

Yes, it generates perfect samples from random noise

Answer

Yes, but only for text data

Answer

No, because it has no decoder

Question 22

In the loss of a variational autoencoder, what two terms are combined?

Accepted Answer

Reconstruction loss and KL-divergence regularization

Answer

Cross-entropy and hinge loss

Answer

Accuracy and precision

Answer

Two separate reconstruction losses

Question 23

What happens if the bottleneck of an autoencoder is made too small?

Accepted Answer

Too little information is kept, causing high reconstruction error (underfitting)

Answer

The model always reconstructs perfectly

Answer

Training becomes impossible to start

Answer

It automatically becomes a VAE

Question 24

What is a contractive autoencoder?

Accepted Answer

One that penalizes the Jacobian to make the code robust to small input changes

Answer

One that shrinks the dataset before training

Answer

One that uses only contraction layers

Answer

One that has no loss function

Question 25

Why are the encoder and decoder typically trained together rather than separately?

Accepted Answer

So both learn compatible mappings via one shared reconstruction loss

Answer

Because separate training is computationally impossible

Answer

To avoid using the latent code

Answer

Because only the decoder needs gradients

Question 26

How can the latent code from a trained autoencoder be used in transfer learning or downstream tasks?

Accepted Answer

The encoder is reused as a feature extractor for downstream models

Answer

The decoder is used to label data

Answer

Only the loss value is transferred

Answer

The latent code cannot be reused

Question 27

In an image denoising autoencoder, what are the input and the target during training?

Accepted Answer

Input is the noisy image; target is the clean original image

Answer

Input is the clean image; target is the noisy image

Answer

Both input and target are random noise

Answer

Input is a label; target is an image

Question 28

What does it mean for an autoencoder's latent space to be 'smooth' or 'continuous', and why is it desirable?

Accepted Answer

Nearby latent points decode to similar meaningful outputs, enabling interpolation

Answer

The latent space has no neurons

Answer

All latent values are exactly zero

Answer

The decoder ignores the latent space

Question 29

Which framework code snippet correctly defines a simple undercomplete autoencoder in Keras?

Accepted Answer

Encoder Dense(32) bottleneck then Dense(784) output, fit with x_train as both x and y

Answer

A single Dense(784) layer trained with class labels

Answer

Two encoders and no decoder, fit on random labels

Answer

Only a softmax classification layer

Question 30

Why is overfitting still a concern for autoencoders, and how can it be reduced?

Accepted Answer

It can memorize data; reduce it with noise, sparsity, regularization, or smaller bottleneck

Answer

Autoencoders can never overfit

Answer

Overfitting is fixed only by adding more layers

Answer

It is solved by removing the decoder

Test yourself