What is deep learning?
While sometimes seemingly used interchangeably, the subtle differences in meanings between artificial intelligence (AI), machine learning and deep learning may cause some confusion. In contrast to natural intelligence displayed by humans and other animals, AI refers to machines mimicking human cognitive functions such as problem solving or learning. So when a machine understands human speech or can compete with humans in a game of chess, AI applies.
Machine learning is the current state-of-the-art application of AI and is largely responsible for its recent rapid growth. Based on the idea of giving machines access to data so that they can learn for themselves, machine learning has been enabled by the internet, and the associated rise in digital information being generated, stored and made available for analysis. Building on AI concepts, machine learning focuses on solving real-world problems through architectures such as artificial neural networks designed to imitate human decision making.
Deep learning concentrates on a subset of machine-learning techniques, with the term "deep" generally referring to the number of hidden layers in the deep neural network. While a conventional neural network may contain a few hidden layers, a deep network may have tens or hundreds of layers. In deep learning, a computer model learns to perform classification tasks directly from text, sound or image data. In the case of images, deep learning requires substantial computing power and involves feeding large amounts of labeled data through a multi-layer neural network architecture to create a model that can classify the objects contained within the image.
Major technology companies such as Google, Facebook, IBM, Intel and Microsoft have invested in deep learning for some time, but more recently, machine-vision software companies have begun to apply deep learning within their products, or base their entire product on it. Machine vision has predominantly relied on rules-based machine-vision algorithms that excel in applications where you know exactly what you're looking for. Classic edge detection, blob, object- and feature-location algorithms generally excel in tasks requiring sub-pixel accuracy for precision measurements or robot guidance.
The value of deep learning in machine-vision applications stems from its ability to make human-like judgments of part quality and other example-based decisions. Verifying the presense of bolts, brackets, foam pads and straps on car seat assemblies, for example, can challenge traditional machine vision systems if subcomponents come from a variety of suppliers with variations in color and texture. In such applications, deep learning helps machine-vision systems cope with the range of acceptable part appearances. Likewise, our cover story this month on page 15 discusses how a single, universal, pre-trained classifier based on deep-learning algorithms enables identification of a wide range of typefaces in optical character recognition applications. Also, on page 7 we cover development of a neural network designed to identify six different defect classes on reflective metallic surface images. As always, I hope you enjoy this issue.
John Lewis, Editor in Chief