Data Augmentation (Artificial Data Expansion)¶

Why It Matters¶

Data collection is expensive and slow; data augmentation is cheap and fast. The key insight: not all variation matters. If you can identify irrelevant variation (horizontal flips for image recognition, synonyms for text), you can generate new training examples that teach the model to ignore these nuisances.

The Key Move¶

Identify transformations that change the input but not the underlying pattern you care about. Apply these transformations to generate additional training examples. Critically: the transformations must preserve label correctness. For images: rotation, flipping, cropping, color jittering. For text: synonym replacement, back-translation. For time series: jittering, scaling, warping.

Classic Application¶

Computer vision. Training a digit classifier with only 100 examples per digit. Augment by rotating ±15 degrees, shifting by a few pixels, scaling slightly. Effective training set expands from 1,000 to 10,000+ examples, dramatically improving generalization. The model learns "digit identity is invariant to these transformations."

Surprising Application¶

Deliberate practice and skill acquisition. You have limited access to real practice scenarios (expensive, rare, dangerous). Create augmented practice: simulations, role-playing, varied contexts. A negotiator practices the same core skill (identifying interests) across synthetic scenarios (salary, vendor contracts, family disputes). Augmentation builds robustness to surface variation.

Failure Modes¶

Invalid augmentations that change the label - rotating a "6" by 180 degrees makes it a "9," not an augmented "6." Over-augmentation creating unrealistic examples that hurt generalization. Augmentation doesn't replace real data diversity - it only captures the variations you can specify.

Go Deeper¶

Shorten & Khoshgoftaar (2019), "A Survey on Image Data Augmentation for Deep Learning," J Big Data; Cubuk et al. (2019), "AutoAugment: Learning Augmentation Strategies," CVPR