Data Augmentation

"In 2012, AlexNet shocked ImageNet largely thanks to one trick: random crops and horizontal flips. A decade later, every state-of-the-art vision model still depends on augmentation, and language models lean on back-translation, paraphrasing, and Mixup. Augmentation is a free regularizer when you do it right — and a beautifully invisible bug when you augment the test set or invent labels that don't exist in the real world."

Learning Objectives

After this lesson, you will be able to:

Use image tricks (rotate, flip, crop, change brightness, Mixup, CutMix) to make your training set bigger and your model smarter without introducing incorrect labels
Use text tricks (swap synonyms, translate back-and-forth, insert context words) to multiply your NLP training examples
Apply SMOTE to generate synthetic minority samples for imbalanced tabular datasets by interpolating between existing nearest neighbors
Know when to augment vs. just collect more real data, and avoid augmentation mistakes that teach your model the wrong thing — always augment after splitting, never before

Prerequisites

Before You Start

Review

Why

Why This Matters

How self-driving cars train on millions of simulated scenarios -- real crash data is rare and dangerous to collect, so companies like Tesla and Waymo use data augmentation to generate millions of synthetic driving scenarios: rain on a sunny road, a pedestrian shifted 2 feet left, a stop sign partially blocked by snow; each augmented scene teaches the AI to handle situations it has never seen in real life
How your phone recognizes your face from any angle -- Face ID was not trained on billions of faces; Apple used augmentation to rotate, flip, adjust lighting, and warp a smaller set of images, teaching the model to recognize you whether you are in bed, at your desk, or squinting in bright sunlight
How TikTok and Instagram train content filters with limited examples -- harmful content is rare, so there are very few training examples of each violation; augmentation (rephrasing text, cropping images, adding noise) multiplies the data so the filter catches bad content even when it has only seen a handful of real examples
Build this --> Build an augmented dataset: take 10 photos of your pet (or any object), apply 5 augmentations to each (rotate, flip, crop, brightness, blur) to create 50 training images, then compare how a simple classifier performs with the original 10 vs. the augmented 50

Data Augmentation

After this lesson, you will be able to:

Before You Start

Your Reflection

Continue Learning

Discussion

#100 Angles on the Same Subject

#Image Augmentation

#Geometric Transformations

#Photometric Transformations

#Advanced Image Augmentation

#Text Augmentation

#Simple Text Augmentation

#Advanced Text Augmentation

#Tabular Data Augmentation

#SMOTE (Synthetic Minority Over-Sampling Technique)

Limitations of SMOTE

#Other Tabular Methods

#When to Augment vs. Collect More Data

#Augment When...

#Collect When...

#The Hierarchy

#When NOT to Augment

#Try It Yourself

#Augmentation in Practice

#How Much Augmentation?

#Augmentation as Regularization

#Key Takeaways

#Quick Check

Deep Dive

Related Lessons