Case study
Several benchmarks exist for image classification. We will use the MNIST image database for this case study. When we used MNIST in Chapter 3, Unsupervised Machine Learning Techniques with clustering and outlier detection techniques, each pixel was considered a feature. In addition to learning from the pixel values as in previous experiments, with deep learning techniques we will also be learning new features from the structure of the training dataset. The deep learning algorithms will be trained on 60,000 images and tested on a 10,000-image test dataset.
Tools and software
In this chapter, we introduce the open-source Java framework for deep learning called DeepLearning4J (DL4J). DL4J has libraries implementing a host of deep learning techniques and they can be used on distributed CPUs and GPUs.
DeepLearning4J: https://deeplearning4j.org/index.html
We will illustrate the use of some DL4J libraries in learning from the MNIST training images and apply the learned models to classify...