Data augmentation quiz

Question 1

What is the primary goal of data augmentation in machine learning?

Accepted Answer

To increase dataset diversity and reduce overfitting

Answer

To reduce the model's number of parameters

Answer

To speed up the training loop

Answer

To normalize pixel values to [0, 1]

Question 2

Which of the following is a label-preserving augmentation for image classification?

Accepted Answer

Horizontally flipping a cat image

Answer

Randomly reassigning labels to images

Answer

Removing 50% of training samples

Answer

Replacing all images with Gaussian noise

Question 3

What does the following PyTorch snippet do to a training image?

Accepted Answer

Applies horizontal flip, random rotation, then converts to tensor

Answer

Resizes all images to 224×224

Answer

Converts to tensor then normalizes with ImageNet mean/std

Answer

Applies random crop and color jitter only

Question 4

Why is horizontal flipping generally safe for CIFAR-10 but potentially harmful for the MNIST digit dataset?

Accepted Answer

Flipping certain digits like 6 or 9 can produce a different valid digit, corrupting labels

Answer

CIFAR-10 images are larger, so flipping is computationally cheaper

Answer

MNIST has more classes, making flipping less risky

Answer

CIFAR-10 uses grayscale images that are flip-invariant

Question 5

What is the difference between online (on-the-fly) and offline data augmentation?

Accepted Answer

Online applies transforms at batch-load time each epoch; offline pre-generates and stores augmented data

Answer

Online augmentation saves augmented data to disk; offline does not

Answer

Offline augmentation applies transforms randomly each epoch; online is fixed

Answer

There is no practical difference between the two approaches

Question 6

SMOTE is a data augmentation technique designed for which kind of problem?

Accepted Answer

Tabular datasets with class imbalance, by synthesizing new minority samples

Answer

Image datasets where images are too large

Answer

Text datasets lacking vocabulary diversity

Answer

Time-series datasets with missing timestamps

Question 7

In Mixup augmentation, two training samples (x1, y1) and (x2, y2) are combined as x̃ = λx1 + (1−λ)x2 and ỹ = λy1 + (1−λ)y2. What does the mixed label ỹ represent?

Accepted Answer

A weighted combination of the two one-hot labels, proportional to λ

Answer

The label of whichever sample had the larger λ weight, chosen deterministically

Answer

Always the label of the first sample x1

Answer

A hard one-hot label chosen randomly from the two classes

Question 8

What does CutOut augmentation do, and why does it help model robustness?

Accepted Answer

Masks a random square region to zero, making the model robust to occlusions

Answer

Cuts the image into patches and shuffles them randomly

Answer

Crops the image to a smaller random sub-region

Answer

Removes the background and keeps only the foreground object

Question 9

What is the key difference between CutOut and CutMix augmentation?

Accepted Answer

CutOut zeros a patch; CutMix replaces that patch with content from another image and blends labels

Answer

CutOut uses color jitter; CutMix uses rotation

Answer

CutMix zeros a patch; CutOut replaces the patch with noise

Answer

They are identical except CutMix operates on text, not images

Question 10

AutoAugment learns augmentation policies automatically. What is the search space it explores?

Accepted Answer

Operations, probability of application, and magnitude of each operation

Answer

Only the probability of applying each fixed transformation

Answer

Only the learning rate and batch size for data loading

Answer

The number of augmented copies to generate per image

Question 11

What is Test-Time Augmentation (TTA) and how does it improve predictions?

Accepted Answer

Averaging predictions over multiple augmented views of each test image

Answer

Augmenting the test set and retraining the model on it

Answer

Applying augmentations only during the final epoch of training

Answer

Reducing test set size by removing augmented duplicates

Question 12

Color jitter augmentation randomly perturbs brightness, contrast, saturation, and hue. Why is hue jitter applied with a much smaller range than brightness jitter in practice?

Accepted Answer

Large hue shifts can fundamentally change object identity, while brightness variation is naturally wide

Answer

Hue values are stored as floats and are numerically unstable with large shifts

Answer

Hue jitter is more computationally expensive and must be kept small

Answer

Brightness jitter is applied after hue jitter, resetting its effect

Question 13

What Python library function applies elastic deformation to images, and why is it especially useful in medical imaging?

Accepted Answer

ElasticTransform from albumentations; useful because it simulates realistic organ/tissue shape variation

Answer

RandomCrop from torchvision; useful because medical images are always too large

Answer

GaussianBlur from PIL; useful because MRI images are always sharp

Answer

RandomFlip from keras; useful because medical images are always symmetric

Question 14

For an object detection task, why must bounding box coordinates be transformed alongside the image during augmentation?

Accepted Answer

Bounding box coordinates must be transformed with the same geometric parameters as the image

Answer

Bounding boxes are pixel-independent and never need to be updated

Answer

Only the image is augmented; bounding boxes remain fixed to the top-left corner

Answer

Bounding boxes are re-annotated manually after each augmentation batch

Question 15

What is the Mosaic augmentation technique used in YOLOv4/v5, and what advantage does it provide?

Accepted Answer

Tiles four images into a 2×2 composite with merged bounding box annotations

Answer

Splits one image into four patches and trains on each patch separately

Answer

Randomly selects one of four color palettes and applies it to the image

Answer

Creates a 4-frame video sequence from a single static image

Question 16

In NLP, what is back-translation augmentation and what type of dataset is it best suited for?

Accepted Answer

Translating to an intermediate language and back to generate paraphrases

Answer

Translating training data to English and discarding the original language

Answer

Using a thesaurus to replace every noun with a synonym

Answer

Splitting long sentences into shorter ones to double the dataset size

Question 17

What are the four simple text augmentation operations introduced in the EDA (Easy Data Augmentation) paper?

Accepted Answer

Synonym replacement, random insertion, random swap, random deletion

Answer

Tokenize, lowercase, remove stopwords, and stem

Answer

Translate, back-translate, paraphrase, and summarize

Answer

Sentence splitting, merging, reordering, and masking

Question 18

For time-series data (e.g., ECG signals), which of the following augmentation techniques preserves temporal ordering and is therefore valid?

Accepted Answer

Adding small Gaussian noise to each time step

Answer

Randomly shuffling all time steps in the sequence

Answer

Reversing the entire time series to train on backward sequences

Answer

Replacing every other time step with a zero

Question 19

What is SpecAugment, and for which data modality was it designed?

Accepted Answer

An audio augmentation method that masks frequency and time bands in spectrograms

Answer

A text augmentation method that masks random tokens in a sentence

Answer

An image augmentation method that applies frequency-domain filtering

Answer

A tabular augmentation method that adds noise to spectral features

Question 20

How does using a GAN (Generative Adversarial Network) for data augmentation differ from classical augmentation techniques?

Accepted Answer

GANs generate entirely new synthetic samples by learning the data distribution, not by transforming existing images

Answer

GANs apply the same geometric transforms as classical methods but faster

Answer

Classical augmentation uses deep learning while GANs use hand-crafted rules

Answer

GANs are only used for text augmentation, not image augmentation

Question 21

What is RandAugment, and how does it simplify the AutoAugment search problem?

Accepted Answer

RandAugment reduces the search to two global hyperparameters N and M, removing RL-based policy search

Answer

RandAugment uses reinforcement learning with a smaller reward signal

Answer

RandAugment adds random Gaussian noise and nothing else

Answer

RandAugment randomly selects a pre-computed AutoAugment policy at each epoch

Question 22

Why is data augmentation considered a form of regularization?

Accepted Answer

Augmentation prevents memorization by exposing the model to diverse perturbed versions of each sample

Answer

Augmentation adds L2 penalty terms directly to the loss function

Answer

Augmentation is not related to regularization; they are independent techniques

Answer

Augmentation regularizes by reducing the learning rate proportionally to dataset size

Question 23

What does the `p` parameter in Albumentations transforms control, and why is it important to tune?

Accepted Answer

The probability that the transform is applied to each image

Answer

The number of pixels to shift the image by

Answer

The padding added around the image before transformation

Answer

The precision (float32 vs float16) used for the transform

Question 24

Feature space augmentation (e.g., Mixup applied to embeddings rather than raw images) has what advantage over pixel-space augmentation?

Accepted Answer

Blending in feature space produces more semantically valid combinations that lie on the learned manifold

Answer

Feature space augmentation is faster because embeddings have fewer dimensions than images

Answer

Feature space Mixup always outperforms pixel Mixup on every benchmark

Answer

Feature space augmentation eliminates the need for a labeled dataset entirely

Question 25

What is the correct Keras/TensorFlow way to add data augmentation as layers inside a model so it is applied only during training?

Accepted Answer

Keras augmentation layers apply transforms during training and pass inputs unchanged during inference

Answer

Keras augmentation layers always apply transforms at both train and test time

Answer

Keras augmentation layers are only applied at test time to improve predictions

Answer

Keras augmentation layers must be defined outside the model and applied in a data generator

Question 26

When should you NOT apply data augmentation to a dataset?

Accepted Answer

When transformations change the semantic meaning/label of the data

Answer

When the dataset has fewer than 100 samples

Answer

When using GPU training, since augmentation only runs on CPU

Answer

When the model has more than 10 million parameters

Question 27

How does random erasing differ from CutOut in its implementation?

Accepted Answer

Random erasing fills with random pixel values and varies aspect ratio; CutOut always uses zero-fill fixed square

Answer

Random erasing fills the region with zeros; CutOut fills with random noise

Answer

They are identical; random erasing is just another name for CutOut

Answer

Random erasing is used for text; CutOut is used for images

Question 28

In a class-imbalanced dataset with 1000 majority-class and 50 minority-class samples, which augmentation strategy is most targeted?

Accepted Answer

Apply augmentation only to minority-class samples to balance class frequencies

Answer

Augment all classes equally to keep the imbalance ratio constant

Answer

Augment only the majority class to create more training variety

Answer

Remove majority-class samples until both classes have 50 samples

Question 29

What is the key idea behind Copy-Paste augmentation for instance segmentation?

Accepted Answer

Cuts object instances from one image and pastes them onto another, with updated masks

Answer

Copies model weights from a pretrained network and pastes them into a new model

Answer

Pastes random noise patches onto images to simulate occlusions

Answer

Copies training images and pastes them with shifted color histograms

Question 30

What is the difference between geometric and photometric (appearance) augmentations? Give one example of each.

Accepted Answer

Geometric augmentations change spatial layout (e.g., rotation); photometric ones change pixel intensities (e.g., brightness)

Answer

Geometric augmentations change pixel colors; photometric augmentations change spatial layout

Answer

Both change only pixel values, not positions

Answer

Geometric augmentations require GPU; photometric ones run on CPU only

Test yourself