Data Augmentation Vocabulary Exercises

1 / 5

A computer vision engineer says: "We applied flip, crop, and rotate augmentations during training to reduce overfitting."

What is the primary purpose of these geometric image augmentations?

Geometric augmentations — vocabulary:

Augmentation	What it does	Label preserved?
Horizontal flip	Mirrors image left-right	Yes (usually)
Random crop	Extracts a random subregion	Yes
Rotation	Rotates by a random angle	Yes (small angles)
Colour jitter	Randomly changes brightness, contrast, saturation	Yes

Key vocabulary: label-preserving transformation, overfitting, generalisation, training set diversity, on-the-fly augmentation.

2 / 5

An NLP engineer describes their pipeline: "We use back-translation as a text augmentation technique to generate additional training examples for our low-resource sentiment classifier."

What is back-translation in the context of text augmentation?

Back-translation — text augmentation:

Step	Example
Original (EN)	"The product quality is excellent."
Translate to DE	"Die Produktqualitat ist hervorragend."
Translate back to EN	"The quality of the product is outstanding."

Other text augmentation techniques: synonym replacement, random insertion, random deletion, sentence shuffling, paraphrasing with an LLM. Key vocabulary: back-translation, paraphrase generation, pivot language, low-resource NLP.

3 / 5

A data science team defines their augmentation strategy for a medical image classification task.

Which statement best describes what an "augmentation strategy" involves?

Augmentation strategy — key design decisions:

Decision	Considerations
Augmentation types	Which transforms are label-preserving for this task?
Magnitude/parameters	How aggressive? (e.g., rotation +-10 vs. +-90 degrees)
Application probability	How often is each augmentation applied? (e.g., 50% of samples)
Domain alignment	Do augmentations reflect real deployment conditions?

Key vocabulary: augmentation strategy, augmentation policy, AutoAugment, RandAugment, label-preserving, domain-specific augmentation.

4 / 5

A ML engineer describes their training setup: "We built an augmentation pipeline using Albumentations that applies a random sequence of transforms on the fly during each training epoch."

What does "on-the-fly" augmentation in a pipeline mean?

Augmentation pipeline modes:

Mode	When applied	Storage	Diversity
On-the-fly (online)	Per batch at training time	No extra storage	Maximum (new transform each epoch)
Offline (pre-generated)	Before training; saved to disk	Fixed multiplier (e.g., 5x dataset)	Limited (fixed transforms)

Tools: Albumentations, torchvision.transforms, tf.data, imgaug. Key vocabulary: augmentation pipeline, online augmentation, data loader, transform composition.

5 / 5

A data scientist applies SMOTE (Synthetic Minority Over-sampling Technique) to address class imbalance in a tabular dataset.

How does SMOTE generate synthetic samples?

SMOTE — interpolation-based augmentation:

Step	Description
1. Select minority sample	Pick a random minority-class record
2. Find k nearest neighbours	Identify k closest minority-class neighbours in feature space
3. Interpolate	Generate a new sample along the line between the sample and a chosen neighbour

Variants: Borderline-SMOTE, ADASYN, SMOTE-ENN. Key vocabulary: class imbalance, over-sampling, under-sampling, feature interpolation, minority class, imbalanced learning.