Data augmentation

What it is

Data augmentation is a way to make more training data from data that already exists. In machine learning, a training dataset is the set of examples a model learns from. Augmentation can change an image by rotating, flipping, translating, or adjusting brightness, as long as the change does not alter the label or core meaning. In one example, a cat image can be flipped or rotated and still count as a cat. The goal is to improve model performance and robustness, especially when real data are limited or classes are unevenly represented.

[1] Data Augmentation and Synthetic Data Generation in Rare ... [2] Conditional Data Synthesis Augmentation - PMC - NIH

How it works

The main idea is to expose a model to more variety without collecting new real samples. Classical augmentation uses geometric changes, which alter spatial layout, and photometric changes, which alter visual properties like brightness or contrast. Some systems do this offline, meaning new samples are created before training and stored. Others do it online, meaning new samples are created during training. Online methods can use a generative adversarial network (GAN), which is a model that creates new samples, but GAN training can be unstable. More advanced approaches also use diffusion models, which add noise and then train a neural network to remove it.

[1] Data Augmentation and Synthetic Data Generation in Rare ... [2] Conditional Data Synthesis Augmentation - PMC - NIH [3] A solid waste classification using reinforcement learning boosted by ...

Why it helps

Augmentation can reduce overfitting, which happens when a model learns the training examples too narrowly and performs poorly on new data. By showing altered versions of the same underlying example, the model learns features that are less tied to one exact image or one exact sample. This can help with small sample size, class imbalance, and limited variability. In rare disease research, it may also support dataset expansion and make models more robust, but the benefits still depend on whether the altered data remain biologically plausible.

[1] Data Augmentation and Synthetic Data Generation in Rare ... [3] A solid waste classification using reinforcement learning boosted by ...

Limitations

Augmentation cannot create truly novel patient types from nothing; it only changes what is already present. If the original dataset is biased or incomplete, augmentation can spread those same problems to more samples. It may also fail when a transformation changes the meaning of the data. In addition, some generated samples can contain artifacts, and validation is still necessary to check that the augmented data are realistic and useful. The literature also notes inconsistent reporting of dataset size before and after augmentation, which makes comparisons harder.

[1] Data Augmentation and Synthetic Data Generation in Rare ... [3] A solid waste classification using reinforcement learning boosted by ...

Sources and further reading

This general lesson was generated from public source material. It explains a concept; it does not interpret a private paper or provide medical advice.

1. Data Augmentation and Synthetic Data Generation in Rare ...
2. Conditional Data Synthesis Augmentation - PMC - NIH
3. A solid waste classification using reinforcement learning boosted by ...