More Complex Problems
Classification, regression, and clustering algorithms are sometimes referred to as shallow learning, in contrast to deep learning. Admittedly, the distinction between shallow learning and deep learning is a bit sketchy and cursory; yet, it marks the point of separating problems that can be solved with a relatively straight algorithm from those that require the introduction of some flavor of neural networks (more or less deep in terms of constituent layers) or the pipelining of multiple straight algorithms. Typically, these problems revolve around the area of cognition such as computer vision, creative work, and speech synthesis.
Image processing began in the late 1960s when a group of NASA scientists had the problem of converting analogic signals to digital images. The core of image processing is the simple application of mathematical functions to a matrix of pixels. A much more enhanced form of image processing is computer vision.
Computer vision isn’t limited to processing data points but attempts to recognize patterns of pixels and how they match to forms (objects, animals, persons) in the real world. Computer vision is the branch of machine learning devoted to the emulation of the human eye, capable of capturing images and recognizing and classifying them based on properties such as size, color, and luminosity.
In the realm of computer vision, image classification is one of the most interesting sectors, especially for its applications to sensitive fields such as health care and security. Image classification is the process of taking a picture (or a video frame), analyzing it, and producing a response in the form of a categorical value (it’s a dog) or a set of probabilistic values (70 percent, it’s a dog; 20 percent, it’s a wolf; 10 percent, it’s a fox). In much the same way, an image classifier can guess mood, attitude, or even pain.
Even though many existing cloud services can recognize and classify images (even video frames), the problem of image classification can hardly be tackled outside a specific business context. In other words, you can hardly take a generic public cloud cognitive service and use it to process medical images (of a certain type) or monitor the live stream of a public camera. You need specific training for the algorithm tailor-made for the scenario you’re facing.
An image classifier is typically a convolutional multilayer neural network. In such a software environment, each processing node receives input from the previous layers and passes processed data to the next. Depending on the number (and type) of layers, the resulting algorithm proves able (or not so able) to do certain things.
A side aspect of computer vision, tightly related to image classification, is object detection. With image classification, you can rely on a class of algorithms capable of looking at live streams of pictures and recognize elements in it. In other words, image classification can tell you what is in the processed picture. Object detection goes one step further and operates a sort of multiclass classification of the picture, telling about all the forms recognized and also about their relative position.
Object detection is very hot in technologies like self-driving cars and robotics. Advanced forms of object detection can also identify bounding boxes for the form to find and even draw precise boundaries around it. Object detection algorithms typically belong to either of two classes—classification-based or regression-based.
In this context, classification and regression don’t refer to the straight shallow learning algorithms covered earlier in the chapter but relate to the learning approach taken by the neural network to come to a conclusion.
Text analytics consists of parsing and tokenizing text, looking for patterns and trends. It is about learning relationships between named entities, performing lexical analysis, calculating and evaluating the frequency of words, and identifying sentence boundaries and lemmas. In a way, it’s a statistical exercise of data mining and predictive analysis applied to text with the ultimate goal of taking software to interact with humans using the same natural language.
A typical application of text analytics is summarizing, indexing, and tagging the content of large digital free text databases and documents such as the comments (and complaints) left by customers of a public service. Text analytics often goes under the more expressive name of natural language processing (NLP) and is currently explored in more ambitious scenarios such as processing a live stream, performing speech recognition, and using recognized text for further parsing and information retrieval. Natural language processing applications are commonly built on top of neural networks in which the input text passes through multiple layers to be progressively parsed and tokenized until the networks produce a set of probabilistic intents.
There are quite a few applications of NLP available in the industry, buried in the folds of enterprise frameworks used in answering machine applications and call centers. However, if you just want to explore the power of the raw NLP, research a few of the existing test platforms, such as https://knowledge-studio-demo.ng.bluemix.net. The tool parses text, an excerpt of a police car accident report, and automatically extracts relevant facts, such as age of the involved people, characteristics of involved vehicles, location, and time.