Back to Subject

Unit 4: Autoencoders and Pre-trained CNN - Practice Quiz

INT422 — Deep Learning 60 Questions
0 Correct 0 Wrong 60 Left
0/60

1 What is the primary goal of an autoencoder?

Introduction to autoencoders Easy
A. To classify labeled samples
B. To sort input records
C. To reconstruct its input
D. To predict future values

2 What type of neural network is an autoencoder?

Introduction to autoencoders Easy
A. A decision tree
B. A reconstruction network
C. A rule-based system
D. A clustering algorithm

3 In a basic autoencoder, the desired output is usually what?

Introduction to autoencoders Easy
A. A class label
B. The original input
C. A random vector
D. A cluster number

4 Which task commonly uses autoencoders?

Need for autoencoders Easy
A. A detailed manual assignment of class names to every sample before the model begins training
B. Database indexing
C. Dimensionality reduction
D. Program compilation

5 Why are autoencoders useful for feature learning?

Need for autoencoders Easy
A. They guarantee perfect classification
B. They learn useful representations
C. They remove all input data
D. They manually label samples

6 How can an autoencoder help remove noise from data?

Need for autoencoders Easy
A. By increasing the noise level
B. By converting the input into manually defined labels using a fixed set of human-written rules
C. By reconstructing a clean version
D. By deleting every feature

7 Which component maps the input to a latent representation?

Architecture of autoencoder Easy
A. Optimizer
B. Classifier
C. Decoder
D. Encoder

8 Which component reconstructs data from the latent representation?

Architecture of autoencoder Easy
A. Decoder
B. Encoder
C. Labeler
D. Sampler

9 What is the compact middle representation of an autoencoder commonly called?

Architecture of autoencoder Easy
A. Latent code
B. Output target
C. Class label
D. A complete collection of all training examples stored without any transformation or dimensional reduction

10 Which loss is commonly used to measure reconstruction error for continuous-valued inputs?

Architecture of autoencoder Easy
A. Recall score
B. Classification accuracy
C. Precision score
D. Mean squared error

11 Which autoencoder is trained to recover clean input from corrupted input?

Types of autoencoders Easy
A. Sparse autoencoder
B. Variational autoencoder
C. Convolutional classifier
D. Denoising autoencoder

12 Which type of autoencoder encourages only a small number of neurons to be active?

Types of autoencoders Easy
A. A network that requires every hidden neuron to remain active at its maximum value for all inputs
B. Sparse autoencoder
C. Basic classifier
D. Denoising autoencoder

13 Which type of autoencoder is especially suitable for image data?

Types of autoencoders Easy
A. Sequence-only decoder
B. Convolutional autoencoder
C. Decision tree
D. Linear regressor

14 What distinguishes an undercomplete autoencoder?

Types of autoencoders Easy
A. Its hidden representation is larger than every input and stores all values without learning useful features
B. Its output contains class labels
C. Its decoder is always removed
D. Its latent space is smaller than its input

15 Which part of an autoencoder stores the compressed representation?

Data compression using autoencoders Easy
A. Loss function
B. Training label
C. Bottleneck layer
D. Output layer

16 During autoencoder-based decompression, what does the decoder receive?

Data compression using autoencoders Easy
A. The loss value
B. The class label
C. The complete original dataset together with all labels and training settings
D. The compressed code

17 What is usually sacrificed when compression greatly reduces the amount of stored information?

Data compression using autoencoders Easy
A. Reconstruction accuracy
B. Number of labels
C. Training order
D. Network name

18 What does a variational autoencoder learn for each latent variable?

Variational autoencoders Easy
A. A sorting rule
B. An exact copy of every training input stored permanently in the hidden layers
C. A probability distribution
D. A fixed class name

19 Which values are commonly predicted by a variational autoencoder's encoder?

Variational autoencoders Easy
A. Label and category
B. Mean and variance
C. Minimum and maximum
D. Precision and recall

20 What is a common use of variational autoencoders?

Variational autoencoders Easy
A. Sorting database rows
B. Assigning manual labels
C. Generating new data samples
D. Replacing all neural network layers with fixed mathematical rules that never learn from data

21 An autoencoder is trained on unlabeled images by using each input image as its target. What type of learning setup does this represent?

Introduction to autoencoders Medium
A. Self-supervised learning
B. Reinforcement learning
C. Adversarial learning
D. Supervised classification

22 An autoencoder produces outputs that are nearly identical to its inputs but learns no useful features. Which design choice most likely caused this behavior?

Introduction to autoencoders Medium
A. A normalized input distribution
B. An overly restrictive latent layer
C. An excessively powerful identity path
D. A moderately sized training set

23 Which loss is most appropriate for reconstructing images whose pixel values are real numbers and approximately Gaussian-distributed?

Introduction to autoencoders Medium
A. Triplet loss
B. Mean squared error
C. Categorical cross-entropy
D. Hinge loss

24 A company has thousands of unlabeled sensor records and wants compact features for a later classification task. Why is an autoencoder suitable?

Need for autoencoders Medium
A. It guarantees perfectly separated classes
B. It learns representations without class labels
C. It removes the need for a classifier
D. It assigns labels through clustering rules

25 An autoencoder is trained only on normal machine-operation data. How can it be used to detect equipment faults?

Need for autoencoders Medium
A. Flag samples with high reconstruction error
B. Flag samples with low reconstruction error
C. Select samples with the highest class score
D. Select samples with the smallest gradients

26 Compared with principal component analysis, what is a key reason to use a nonlinear autoencoder for dimensionality reduction?

Need for autoencoders Medium
A. It always finds orthogonal latent features
B. It guarantees a unique representation
C. It can model nonlinear data manifolds
D. It requires no optimization procedure

27 An autoencoder maps to and then reconstructs . Which component creates ?

Architecture of autoencoder Medium
A. The encoder
B. The loss function
C. The decoder
D. The output target

28 For an autoencoder with encoder and decoder , which expression represents the full reconstruction process?

Architecture of autoencoder Medium
A.
B.
C.
D.

29 A grayscale image has pixel values in . Which output activation is a reasonable choice for the decoder?

Architecture of autoencoder Medium
A. Softmax over pixels
B. Unbounded linear
C. Sigmoid
D. Leaky ReLU

30 An autoencoder has input dimension 64 and latent dimension 128. Which technique can still prevent it from learning a trivial identity mapping?

Architecture of autoencoder Medium
A. Fixing every weight to one
B. Removing the reconstruction loss
C. Duplicating each input feature
D. Applying a sparsity constraint

31 A model receives corrupted images as inputs but is trained to reproduce the original clean images. What type of autoencoder is this?

Types of autoencoders Medium
A. Denoising autoencoder
B. Sparse autoencoder
C. Variational autoencoder
D. Contractive autoencoder

32 Which modification most directly creates a sparse autoencoder?

Types of autoencoders Medium
A. Sampling every latent coordinate
B. Penalizing excessive hidden activations
C. Sharing all encoder and decoder weights
D. Adding noise only to target samples

33 A researcher wants the learned representation to change very little when the input is slightly perturbed. Which autoencoder is most suitable?

Types of autoencoders Medium
A. Sequence autoencoder
B. Vanilla autoencoder
C. Contractive autoencoder
D. Variational autoencoder

34 Which autoencoder architecture is generally most suitable for preserving local spatial patterns in images?

Types of autoencoders Medium
A. Convolutional autoencoder
B. Recurrent sequence autoencoder
C. Fully linear autoencoder
D. Sparse tabular autoencoder

35 A grayscale image is represented by 1024 values. An encoder maps it to 64 latent values. Ignoring metadata and numeric precision, what is the compression ratio?

Data compression using autoencoders Medium
A.
B.
C.
D.

36 An autoencoder compresses data effectively on its training set but reconstructs unseen samples poorly. Which change is most likely to improve practical compression?

Data compression using autoencoders Medium
A. Train until the training error reaches zero
B. Increase capacity without changing the data
C. Store the original samples with each code
D. Use regularization and more representative data

37 Why must a decoder or its equivalent be available when distributing data compressed by an autoencoder?

Data compression using autoencoders Medium
A. The encoder automatically stores all original inputs
B. The reconstruction loss contains the original sample
C. The latent code is always human-readable
D. The latent code requires learned reconstruction rules

38 In a variational autoencoder, the encoder outputs and instead of one fixed code. What do these values describe?

Variational autoencoders Medium
A. A decoder weight distribution
B. A reconstruction-error distribution
C. A set of class probabilities
D. A latent probability distribution

39 Which formula represents the reparameterization trick used to sample a VAE latent variable?

Variational autoencoders Medium
A.
B.
C.
D.

40 What is the main role of the KL-divergence term in the VAE objective?

Variational autoencoders Medium
A. Maximize differences between input samples
B. Convert continuous codes into class labels
C. Force every reconstruction to be identical
D. Align the latent posterior with the prior

41 An autoencoder is trained with input-target pairs . After training, which property most directly distinguishes a useful autoencoder from a trivial identity mapping?

Introduction to autoencoders Hard
A. The output layer uses a different dimension from the input layer
B. The reconstruction loss is evaluated only on the latent code
C. The decoder has more parameters than the encoder
D. The latent representation preserves task-relevant structure under a constraint

42 A dataset contains high-dimensional sensor vectors with strong nonlinear correlations. PCA reconstruction is poor, but increasing the PCA dimension is computationally expensive. Why might an autoencoder provide a better compression model?

Need for autoencoders Hard
A. It guarantees lossless reconstruction at every code dimension
B. It learns nonlinear manifolds through parameterized transformations
C. It estimates correlations without using an objective function
D. It removes the need to select a compression dimension

43 For an autoencoder with encoder and decoder , suppose the latent dimension is larger than the input dimension and no regularization is used. Which failure mode is most likely?

Architecture of autoencoder Hard
A. The model may memorize the training examples
B. The decoder must produce a constant output
C. The reconstruction error must equal the input variance
D. The encoder is forced to become linear

44 An autoencoder uses sigmoid outputs and binary cross-entropy reconstruction loss for images whose pixels are scaled to . A reconstruction has predicted value for a target pixel of . What is the principal training consequence?

Architecture of autoencoder Hard
A. The pixel reverses the sign of every encoder gradient
B. The pixel contributes a large penalty and gradient
C. The pixel contributes no penalty because sigmoid saturates
D. The pixel is ignored because its target is exactly zero

45 A convolutional autoencoder downsamples an image by a factor of three times. The decoder uses three stride- transposed convolutions. If padding and output padding are chosen inconsistently, which issue can occur?

Architecture of autoencoder Hard
A. The latent variables become statistically independent
B. The output spatial dimensions differ from the input
C. The encoder loses all channel information
D. The reconstruction loss becomes invariant to scale

46 A denoising autoencoder receives generated by corrupting , but its target remains the clean sample . What representation is it explicitly encouraged to learn?

Types of autoencoders Hard
A. A representation with maximum latent variance
B. A representation dependent only on pixel ordering
C. A representation that reproduces noise exactly
D. A representation invariant to the selected corruption

47 A sparse autoencoder adds an activation penalty to the reconstruction loss. If is increased substantially while model capacity remains fixed, what trade-off is most likely?

Types of autoencoders Hard
A. Identical codes with guaranteed better generalization
B. Random codes with unchanged reconstruction quality
C. Denser codes with strictly lower reconstruction error
D. Sparser codes with potentially worse reconstruction

48 A contractive autoencoder penalizes the Frobenius norm of the encoder Jacobian . What behavior does this penalty primarily encourage near training samples?

Types of autoencoders Hard
A. Sensitivity to small input perturbations
B. Local insensitivity of the code to input changes
C. Exact orthogonality between encoder layers
D. Uniformly distributed activation magnitudes

49 An undercomplete autoencoder achieves nearly zero training reconstruction error even though its latent dimension is much smaller than the input dimension. Which explanation is most plausible?

Types of autoencoders Hard
A. The latent dimension automatically expands during optimization
B. The reconstruction target has been removed during training
C. The decoder has learned a training-set lookup behavior
D. The bottleneck guarantees a semantically meaningful code

50 An autoencoder maps each -dimensional input to a -dimensional floating-point code. Ignoring model weights and metadata, what is the nominal code-size reduction relative to storing the input using the same precision?

Data compression using autoencoders Hard
A.
B.
C.
D.

51 A compressed latent vector is quantized from 32-bit floating-point values to 8-bit integers before storage. Which additional issue must be considered when evaluating reconstruction quality?

Data compression using autoencoders Hard
A. Quantization guarantees a lower reconstruction error
B. Quantization introduces distortion not present during training
C. Quantization always increases the latent dimension
D. Quantization makes the decoder independent of its weights

52 Two autoencoders use the same latent dimension and achieve the same mean squared reconstruction error. Model A preserves rare but safety-critical events poorly, while Model B preserves them well. Which evaluation is more appropriate for selecting Model B?

Data compression using autoencoders Hard
A. Use the number of decoder parameters alone
B. Use a uniformly weighted pixel error
C. Use only the average latent activation
D. Use an application-weighted distortion measure

53 A decoder reconstructs an image from a code that was produced by an encoder trained only on natural images. The code is used to compress medical images containing structures absent from the training set. What is the main risk?

Data compression using autoencoders Hard
A. The decoder may remove out-of-distribution structures
B. The code will necessarily use more bits than the input
C. The latent representation will have no numerical values
D. The decoder will become a lossless universal compressor

54 An autoencoder has reconstruction loss . If every input is multiplied by a constant but the architecture and learning rate are unchanged, why can optimization behavior change substantially under MSE?

Introduction to autoencoders Hard
A. MSE scales approximately by
B. MSE is converted automatically into cross-entropy
C. MSE changes only when the latent dimension changes
D. MSE becomes independent of the target scale

55 In a VAE, the encoder outputs and , and sampling uses with . Why is this reparameterization useful?

Variational autoencoders Hard
A. It forces the posterior variance to equal one
B. It replaces the reconstruction objective with classification
C. It permits gradients to pass through the sampling operation
D. It removes stochasticity from the latent variable

56 For a VAE with posterior and prior , which term is the KL divergence?

Variational autoencoders Hard
A.
B.
C.
D.

57 A VAE's KL term is weighted by a very large coefficient relative to the reconstruction term. Which outcome is most likely if optimization is stable?

Variational autoencoders Hard
A. The decoder becomes unnecessary for reconstruction
B. The posterior remains close to the standard normal
C. The posterior becomes highly input-specific
D. The latent variables become deterministically unique

58 Suppose a VAE reconstructs training examples well, but samples generated from are implausible. Which diagnosis best explains this result?

Variational autoencoders Hard
A. The decoder has no dependence on latent variables
B. The reconstruction loss must be exactly zero
C. The aggregated encoded distribution may not match the prior
D. The input dimension is necessarily larger than the code

59 A VAE uses binary cross-entropy summed over all pixels and a KL divergence averaged over the batch. If image resolution is doubled in both dimensions without retuning weights, what effect is likely?

Variational autoencoders Hard
A. Both terms remain identically scaled
B. The latent prior changes from Gaussian to uniform
C. The KL term gains relative dominance
D. The reconstruction term gains relative dominance

60 An adversarial autoencoder matches the aggregated latent distribution to a prior using a discriminator, while an ordinary autoencoder uses only reconstruction loss. What does the adversarial constraint primarily provide?

Types of autoencoders Hard
A. A mechanism for shaping the latent distribution
B. A proof that the prior is the true data distribution
C. A replacement for the encoder network
D. A guarantee of pixel-perfect reconstruction
Previous MCQ
Unit 3
Next MCQ
Unit 5