Thread by @Jeande_d

One of the techniques that have accelerated machine learning on insufficient real-world datasets is data augmentation.

Data augmentation is the art of creating artificial data.

For example, you can take an image, flip it, change color, and now you have a new image.

Yes, data augmentation is the art of creating artificial data to expand a given small dataset.

It has shown that it works so well(most of the time), and it remarkably handles overfitting.

Nearly most types of data can be augmented, but I have noticed that it works well in unstructured data(images, video, sound).

So, this thread will focus more on images, sounds, and videos.

For more about structured vs unstructured data 👇

https://twitter.com/Jeande_d/status/1418174185310932994?s=20

But why do we have to augment data?

Let's say that you want to build an image classifier (say for a rare real-world use case) and with a lot of effort, you can only get 50 images.

Doubling the efforts, you can get 20 images more. You would likely want to find scalable ways of boosting your dataset.

Thanks to advances in ML techniques, this is not impossible.

Images are probably the most data that augmenting will always work so well.

Here are various ways to augment images:

◆Flipping images horizontally/vertically or both
◆Rotating the images to a given degree

More image augmentation options:

◆Changing the color and contrast
◆Zooming the image in/out
◆Adding noises
◆And more (this is an art)

The below image summarizes image augmentation well.

Like any other technique, data augmentation is not always a solution to everything.

It will improve the results when it is applied in regard to the error analysis.

Here is an example of what I mean...

Say you have 500 images of cats and 900 images of dogs. Augmenting cats and dogs evenly will not improve the results.

You would want to create more images of cats than dogs.

One more thing about images augmentation:

The augmented images have to be realistic.

You can virtually generate many image scenarios but if those scenarios are not realistic,

It will hurt the model.

Sounds and videos can be augmented too.

Sound can be augmented by adding noise, slowing/speeding the sound, shifting the time, narrowing or widening the sound.

Videos can be augmented as sound.

You can slow or speed them, shift the time, split the video, and apply some custom transformations to individual clips.

For example, some clips can have a reduced brightness, different color, etc.

Data augmentation is the art that can not only be applied to a small dataset but even to a big dataset.

This is because it introduces some diversity in training samples and this is a good thing.

About data augmentation tooling,

◆ For TensorFlow guys, there are ImageDataGenerator which is so simple to use, and (experimental) augmentation layers.

keras.io/api/layers/pre…

◆ And for Pytorch users (and beyond), there is *Ugly.

github.com/facebookresear…

This is the end of the thread that was about the art of data augmentation.

To summarize,

Data augmentation can be used to expand training samples and preventing overfitting as a result.

Nearly most data can be augmented.

◆Images: Flipping, changing color & contrast, adding noise, and zooming in/out.

◆Videos: Adding noise, slowing/speeding, shifting time, and applying custom changes to individual clips.

◆Sounds: Adding noise, narrowing/widening voice, shifting time, speeding/slowing sound

For more about this art, beyond the thread, here is a full article:

jeande.medium.com/a-handy-notes-…

Thank you for reading.

I am regularly writing about machine learning ideas and best practices. I challenge myself to create actionable content, and my goal is to create more.

You can support me by following @Jeande_d,

Leaving a ♡,

Or sharing the post with your friends.