Data Augmentation in Deep Learning

Whenever you build and train a model for a machine learning task, regardless of its being a classification or regression one, your final goal is to make reliable predictions on new, never seen before input data. In other words, you want your model to generalize well on new data. To achieve this goal, you have to prevent your model from being either excessively adjusted to training data (overfitted), or not capable of capturing pattern in data at all (underfitted). Note that the concepts of overfitting and underfitting are strictly related to the notion of bias-variance trade-off. In this article, I'm going to dwell on the problem of overfitting and how to deal with it.