Deep Learning Essentials

Data augmentation

Deep learning requires a large corpus of training data in order to effectively learn, but sometimes, collecting such data can be very expensive and unrealistic. One way to help is to do data augmentation, by artificially inflating the training set with label, preserving transformations. By increasing the sample amount, it can also help with overcoming the overfitting problem:

• Rule of thumb—such transformations or manipulations must be carefully designed, implemented, and tested. Working techniques can be domain-specific and may not be used generally.
• For image data, some simple techniques include re-scaling, random cropping, rotating, horizontal flipping, color jittering, and adding noise. By adding noise, a deep learning ...