Imagine you’re an art teacher, and you’re teaching your student how to paint apples. But you only have one apple for them to practice with. Your student paints it again and again, but there’s a limit to how much they can learn from painting the same apple in the same position, from the same angle.
So, to help your student improve, you rotate the apple, cut it into pieces, or even use a mirror to reflect it. Now your student can see and paint the apple from many different perspectives. They’re dealing with ‘augmented’ data, because you’ve created a richer, more varied set of information from a single apple.
Data Augmentation in machine learning is similar. If we have a limited amount of data to train our model (like the single apple), it’s not enough for the model to understand all possible variations of the data it may encounter. To solve this, we create new data by making small modifications to the existing ones.
For example, if we’re training a model to recognize images of cats, we might flip the images horizontally or vertically, rotate them, zoom in or out, or change the lighting conditions in the images. Now, the model is not only learning from the original images but also from these new, modified images. This helps the model to generalize better and make more accurate predictions when encountering new, unseen data.
So, data augmentation is like creating many different “views” of the same data to help our models learn and understand better.
« Back to Glossary Index