Applied Machine Learning and AI for Engineers

Chapter 8. Deep Learning

Every model in Part I of this book employed classic machine learning algorithms that form the core of ML itself: logistic regression, random forests, and so on. Such models are often referred to as traditional machine learning models to differentiate them from deep-learning models. Recall from Chapter 1 that deep learning is a subset of machine learning that relies primarily on neural networks, and that most of what’s considered AI today is accomplished with deep learning. From recognizing objects in photos to real-time speech translation to using computers to generate art, music, poetry, and photorealistic faces, deep learning allows computers to perform feats that traditional machine learning does not.

I frequently introduce deep learning to software developers by challenging them to devise an algorithmic means for determining whether a photo contains a dog. If they offer a solution, I’ll counter with a dog picture that foils the algorithm. Traditional ML models can partially solve the problem, but when it comes to recognizing objects in images, deep learning represents the state of the art. It’s not terribly difficult to train a neural network to recognize dog pictures, sometimes more accurately than humans. Once you learn how to do that, it’s a small step forward to recognizing defective parts coming off an assembly line or bicycles passing in front of a self-driving car.

Neural networks have been around for decades, but it’s only in the past 10 years ...