Data Augmentation
10x Your Dataset for Free

The sweet spot is 2-5x your original data. 100 images → 200-500 total (including originals). More than 10x usually doesn't help and can hurt performance. Quality over quantity: 200 diverse examples beat 1000 similar ones. Online augmentation (during training) is often better than offline (pre-generating all variations).

During training (online) is usually better! Creates random variations on-the-fly each epoch, so AI never sees identical examples. Saves disk space. Offline (pre-generate) is better for: very slow transformations, when you need exact reproducibility, or for debugging. Most frameworks (PyTorch, TensorFlow) support online augmentation easily.

No! Augmentation supplements, doesn't replace real data. 1000 diverse real images always beat 100 images augmented to 1000. Real data captures genuine variety that augmentation can't replicate. Best strategy: collect maximum real data feasible, then augment for boost. Exception: when real data collection is impossible or unethical (medical data, rare conditions).

Winning combos: Images - Flip + Rotate (±15°) + Zoom (80-120%) + Brightness. Text - Synonym replacement + Back-translation. Audio - Time stretch + Pitch shift + Background noise. Apply 2-4 augmentations per image, not all at once. Test combinations on validation set. Start simple, add complexity only if performance improves.

Works great for: classification, object detection, speech recognition, text classification. Less effective for: precision tasks (medical diagnosis), meaning-sensitive tasks (sentiment analysis), huge datasets (100k+ examples). Rule of thumb: if humans would still recognize the augmented version correctly, it's probably useful. Domain-specific augmentations often work better than generic ones.

Augmenting test/validation data (never do this!), extreme transformations (upside down images, neon colors), over-augmentation (100x from 10 examples), ignoring domain constraints (flipping medical X-rays), not checking outputs visually, using augmentations that change meaning (text sentiment), and applying augmentation that destroys important features (too much blur).

Test on validation set! Train with and without augmentation, compare validation accuracy. Visual inspection: manually review 20-50 augmented samples to ensure they look realistic. Training dynamics: good augmentation reduces overfitting (training loss >> validation loss), bad augmentation increases noise (both losses high). If validation accuracy drops, reduce augmentation intensity.

Start conservative: Rotation ±15°, Zoom 80-120%, Brightness ±20%, Contrast ±15%, Saturation ±10%. Text: Replace 10-30% of words, back-translation with 1-2 intermediate languages. Audio: Speed 0.8-1.2x, Pitch ±2 semitones, Volume ±6dB. Adjust based on your domain - satellite imagery can handle more rotation than portrait photos.

Yes! It's perfect for balancing. If you have 900 cats, 100 dogs → augment dogs 8-9x. This creates balanced training without deleting cat data. Alternative: SMOTE for tabular data, oversampling minorities. Be careful not to create unrealistic variations just for balance. Sometimes collecting more minority class data is better than extreme augmentation.

Albumentations is fastest and most popular (70+ transforms, GPU support). imgaug has more exotic transformations but slower. PyTorch/TensorFlow built-in augmentations are basic but well-integrated. Domain-specific: MONAI for medical images, nlpaug for text, audiomentations for audio. Choose based on your needs: Albumentations for general computer vision, specialized libraries for specific domains.

Weak augmentation = small, subtle changes (±10° rotation, slight brightness). Strong = dramatic changes (±45° rotation, heavy noise, color shifts). Weak augmentation usually works better for most tasks. Strong augmentation can help when data is very limited or domain has high natural variation (satellite imagery, medical scans). AutoAugment/RAug automatically find optimal strong augmentation policies.

Augmentation can make debugging harder because model sees different data each epoch. Solutions: set random seeds for reproducibility, save augmented samples for inspection, use deterministic augmentation for debugging. Some argue augmentation hurts interpretability - model learns more robust but less specific features. Balance interpretability vs performance based on your needs.