Applied Machine Learning and AI for Engineers

Chapter 9. Neural Networks

Machine learning isn’t hard when you have a properly engineered dataset to work with. The reason it’s not hard is libraries such as Scikit-Learn and ML.NET, which reduce complex learning algorithms to a few lines of code. Deep learning isn’t difficult either, thanks to libraries such as the Microsoft Cognitive Toolkit (CNTK), Theano, and PyTorch. But the library  that most of the world has settled on for building neural networks is TensorFlow, an open source framework created by Google that was released under the Apache License 2.0 in 2015.

TensorFlow isn’t limited to building neural networks. It is a framework for performing fast mathematical operations at scale using tensors, which are generalized arrays. Tensors can represent scalar values (0-dimensional tensors), vectors (1D tensors), matrices (2D tensors), and so on. A neural network is basically a workflow for transforming tensors. The three-layer perceptron featured in Chapter 8 takes a 1D tensor containing two values as input, transforms it into a 1D tensor containing three values, and produces a 0D tensor as output. TensorFlow lets you define directed graphs that in turn define how tensors are computed. And unlike Scikit, it supports GPUs.

The learning curve for TensorFlow is rather steep. Another library, named Keras, provides a simplified Python interface to TensorFlow and has emerged as the Scikit of deep learning. Keras is all about neural networks. It began life as a standalone project ...