Augmentation

Data augmentation refers to techniques used to artificially increase the size of a dataset by applying various transformations on the original data. Having a larger training dataset can lead to better generalization.

How Data Augmentation Works

Let's have a look at the types of data that can be augmented, the purpose and implementation of data augmentation techniques.

1. Type of Data:

Image Data:

• Rotations, zooming, flips (horizontal & vertical), color variations, cropping, and more.
• More advanced techniques include cutout, and mixup.

Text Data:

• Back translation (translating a sentence from the original language to another language and then back to the original language), synonym replacement, and sentence shuffling.

Audio Data:

• Changing pitch, speed, or adding noise.

2. Purpose:

1. Enhance Generalization: By exposing the model to various modifications of the original data, the model becomes more robust and can generalize better to new, unseen data.
2. Balance Datasets: Augmentation can be used to balance classes in datasets by artificially increasing the number of examples in underrepresented classes.

3. Implementation:

1. Most machine learning libraries, such as TensorFlow and PyTorch, have built-in functions or modules for data augmentation.
2. Data augmentation is typically applied during the training process. When a training batch is requested, raw data samples are fetched and augmented on-the-fly before being fed into the model.