Trivialaugment
TrivialAugment is the simplest possible automated augmentation strategy that matches or exceeds complex learned policies — applying exactly one randomly selected transformation at a randomly selected magnitude to each training image, with zero hyperparameters to tune, proving the counterintuitive result that the "dumbest" approach to augmentation is as effective as sophisticated search-based methods like AutoAugment.
What Is TrivialAugment?
- Definition: For each training image, randomly select one augmentation from a pool (Rotate, Shear, Brightness, etc.) and apply it at a randomly selected magnitude — that's it. No search, no N parameter, no M parameter, no policy learning.
- The Philosophy: "What is the simplest thing that could possibly work?" The answer turns out to be: randomly do one thing at a random strength.
- The Surprise: This trivial algorithm matches AutoAugment (5,000 GPU hours of search) and RandAugment (grid search over N and M) — suggesting that augmentation diversity matters more than specific combinations or magnitudes.
Algorithm (Complete)
For each training image:
1. Randomly select ONE operation from the pool
2. Randomly select a magnitude (uniform from 0 to max)
3. Apply the operation at that magnitude
Done.That's the entire algorithm. No loops, no parameters, no search.
Comparison of Augmentation Complexity
| Method | Hyperparameters | Search Cost | Algorithm Complexity |
|---|---|---|---|
| No Augmentation | 0 | 0 | None |
| Manual Augmentation | Many (per-transform) | Human time | Hand-tuned |
| AutoAugment | 25 sub-policies × 2 ops × 3 params | 5,000 GPU hours | RL controller + proxy training |
| RandAugment | 2 (N and M) | Grid search | Random selection, fixed magnitude |
| TrivialAugment | 0 | 0 | Single random operation |
Why Zero Hyperparameters Wins
| Insight | Explanation |
|---|---|
| Random magnitude = implicit adaptive | Each image gets a different strength — some mild, some strong, naturally covering the space |
| One operation = maximum diversity | Over a training epoch, every operation appears equally — no bias toward specific transforms |
| No overfitting to augmentation | Learned policies can overfit to the proxy task or validation set |
| No computational waste | Zero search cost means all compute goes to actual training |
Results: Trivial = SOTA
| Dataset | Model | AutoAugment | RandAugment | TrivialAugment |
|---|---|---|---|---|
| CIFAR-10 | WRN-40-2 | 3.70% | 3.60% | 3.40% |
| CIFAR-100 | WRN-40-2 | 18.40% | 18.60% | 18.10% |
| ImageNet | ResNet-50 | 22.40% | 22.40% | 22.10% |
TrivialAugment matches or beats all more complex methods — with zero hyperparameters and zero search cost.
The Broader Lesson
TrivialAugment demonstrates a recurring theme in machine learning: simple methods with good inductive biases often match complex methods. The specific augmentation policy matters less than having diverse augmentations applied consistently during training.
TrivialAugment is the proof that simplicity wins in data augmentation — achieving state-of-the-art results with zero hyperparameters and zero search cost by randomly applying a single transformation at a random strength to each training image, challenging the assumption that complex learned augmentation policies are necessary for strong performance.