Whether a model learns well or not depends on both architecture and data. Earlier, we talked about many techniques for "starting from the model side" - regularization Dropout、 Stop early. Today, let's change direction and start from the data side to talk about Data Augmentation.
##What is data augmentation?
Simply put, it is to create more training samples that look different and have unchanged labels by making reasonable transformations to existing samples without truly increasing data collection costs, so that the model can see more changes and reduce overfitting.
The reason is very intuitive: if the model is trained only on "upright cats", it may remember that "cats are usually in the middle of the graph, facing in a certain direction"; Once we feed it a cat that flips, cuts, and adds noise, it is forced to learn the 'essential characteristics of a cat' rather than a coincidence of its position.
##Image field: the most classic battlefield
Images are the most widely used field for data augmentation, with common methods including:
-* * Geometric transformations * *: Horizontal flipping, random cropping, rotation, scaling, translation;
-* * Color transformation * *: Adjust brightness, contrast, saturation, color tone, or perform grayscale transformation;
-* * Obstruction and erasure * *: Randomly occlude a portion of the area (such as Random Erasing, Cutout), forcing the model not to only focus on the local area;
-* * Mixed class method * *: Overlaying two images proportionally and mixing labels proportionally (such as Mixup, CutMix) is a more powerful enhancement.
##Text and voice can also be enhanced
Data augmentation is not just about serving images:
-* * Text * *: synonym replacement, random deletion/swapping of words, back translation (translating into another language first and then back), etc;
-* * Voice * *: Add background noise, speed change, tone change, and randomly clip time segments.
Note that text enhancement is more "dangerous" than image enhancement, and changing one word may alter semantics or even labels, so caution must be exercised.
##Key principle: Transformation should 'preserve labels'
The most important principle of data augmentation is that after transformation, the labels of the samples must still hold true. Flip a picture of a 'cat' horizontally, and it will still be a cat; But if you flip a "number 6" up and down, it may become "9" and the label is wrong. So the enhancement strategy must be designed in conjunction with specific tasks and cannot be blindly flipped over.
##Cost and trade-off
-Enhancement can make the data seen in a single training more abundant, usually significantly improving generalization, but it may also slow down training convergence;
-The enhancement strategy itself is also a hyperparameter. If the force is too weak, it is useless. If it is too strong, it will "make the data look different" and instead damage the performance;
-When the amount of data is already large, the benefits of enhancement will decrease, but on small datasets, it is often the most cost-effective move.
##Summary
The core idea of data augmentation is to use inexpensive and controllable transformations to extract more information from limited data. It, like regularization and Dropout, is a means of combating overfitting and improving generalization, but one constrains from the model side and the other expands from the data side. The combination of the two often yields better results.
[Reference source]
-PyTorch official documentation torchvision. transforms (commonly used image data augmentation API)
-Comprehensive compilation of industry information that has been publicly released