Home Knowledge Base AutoAugment

AutoAugment is a reinforcement learning approach to automatically discover optimal data augmentation policies for a given dataset — replacing human intuition ("maybe I should rotate by 15° and adjust brightness?") with a learned search that trains thousands of candidate policies and selects the one that maximizes validation accuracy, discovering non-obvious augmentation combinations (like "Shear + Solarize" or "Equalize + Rotate") that consistently outperform hand-designed strategies.

What Is AutoAugment?

AutoAugment Policy Structure

LevelComponentExample
Policy25 sub-policiesThe complete augmentation strategy
Sub-policy2 sequential operations"Shear + Solarize"
OperationTransform type + probability + magnitude"Rotate with p=0.6 and magnitude=7"

Search Process

StepProcessCompute Cost
1. Controller (RNN) proposes policySamples augmentation operationsMinimal
2. Child network trains with proposed policyTrain small proxy model on subsetHours per policy
3. Validation accuracyEvaluate on held-out dataPart of step 2
4. RL reward signalValidation accuracy → controllerController learns which policies work
5. Repeat 15,000+ timesSearch over policy space5,000 GPU hours ⚠️

Discovered Policies (Surprising Results)

DatasetKey Operations FoundSurprise
CIFAR-10Invert, Equalize, ContrastIntensity transforms > geometric transforms
ImageNetPosterize, Solarize, EqualizeColor quantization helps (unexpected)
SVHNInvert, Shear, TranslateStreet numbers benefit from shearing

AutoAugment vs Later Methods

MethodSearch CostHyperparametersPerformanceYear
AutoAugment5,000 GPU hoursPer-dataset policy search requiredState-of-art at release2019
Fast AutoAugment3.5 GPU hoursDensity matching, no RLComparable to AutoAugment2019
RandAugment0 (no search)Just N (ops) and M (magnitude)Comparable, much simpler2020
TrivialAugment0 (no search)Zero hyperparametersEqual or better2021

The Legacy of AutoAugment

AutoAugment is the pioneering work that proved data augmentation policies can be learned rather than hand-designed — demonstrating significant accuracy improvements by searching over augmentation strategies with reinforcement learning, and inspiring simpler successors (RandAugment, TrivialAugment) that achieve comparable results without the expensive search process.

autoaugmentlearnedpolicy

Explore 500+ Semiconductor & AI Topics

From EUV lithography to CUDA optimization — search the full knowledge base or chat with our AI assistant.