NCA-GENM - Data Analysis and Visualization (10% of the exam) - Section 5.1
Enhance datasets through data augmentation.
Describe data augmentation techniques - such as flipping, cropping, colour jitter, and synthetic generation - that artificially expand training datasets. Recognise when augmentation reduces overfitting by increasing effective dataset diversity without collecting new labelled samples.
Data augmentation
Practice question for this objective
lock_openFree sampleData Analysis and Visualizationeasy
Compared with collecting and labelling additional real-world examples, what is the principal advantage that data augmentation offers?
AIt increases training diversity cheaply by transforming data already on hand, avoiding the cost and time of gathering and annotating new samples.check_circle Correct
BIt guarantees that the augmented samples introduce entirely new real-world scenarios absent from the original collected data.
CIt removes the need for any regularisation, because augmented data alone fully eliminates overfitting in every model.
DIt produces a smaller model with fewer parameters, because augmented inputs let the network discard redundant layers during training.
Augmentation's principal advantage is cheaply increasing training diversity from existing labelled data instead of collecting and annotating new samples. Because augmentation transforms data already collected and labelled, it raises diversity at a fraction of the cost of sourcing and annotating fresh real-world examples, though it cannot invent genuinely new phenomena.
Why A is correct: Correct: augmentation reuses existing labelled data through transformations, expanding diversity at low cost and effort relative to sourcing and annotating new real examples.
Why B is wrong: Tempting because augmentation does add variety, but it derives variants from existing samples and cannot introduce genuinely novel phenomena that were never captured; only new real data can do that.
Why C is wrong: Tempting because augmentation does reduce overfitting, but it complements rather than replaces other regularisers and does not eliminate overfitting universally; the absolute claim is wrong.
Why D is wrong: Tempting because efficiency is desirable, but augmentation affects the data fed to the model, not the model's parameter count or architecture; it does not shrink the network.
Answers that look right on this material and are not. Each one is a distractor from a different question in the NCA-GENM bank for this domain.
It compresses the raw dataset on disk so that fewer bytes are stored while the same number of distinct training samples is retained.
errorWhy it is wrong: Tempting because augmentation pipelines are sometimes applied at load time and seem storage-related, but augmentation increases the variety of training signal rather than reducing storage, and it does not compress the original files.
Normalising amplitude levels across all recordings to a fixed decibel target
errorWhy it is wrong: Amplitude normalisation standardises volume but produces no new temporal variations, so it does not expand the diversity of speech rates seen during training.
Generating synthetic defect images with a diffusion model trained on the same dataset
errorWhy it is wrong: Diffusion-based synthesis can introduce label noise and distribution shift because the generative model itself may not have seen enough defect examples, making this riskier than deterministic transforms.