Machine Learning

Machine Learning can be approached from two different perspectives: 1) The type of application, or 2) The Learning paradigm.

Segmentation of Machine Learning by type of application:

The tasks performed by Machine Learning can be grouped into typical applications. Topping the list of usage is classification. Most of image processing or computer vision is based on classification, from automatic friend tagging on Facebook to tumor detection on an MRI, from quality control on a manufacturing line to obstacle identification by autonomous vehicles.

The second most popular application is continuous estimation. This application is also known by the class of algorithm that it implements: regression. Determining the likely price for of a house or the yearly sales for a product, forecasting demand for electricity or the number of years that an employee might stay in a given position are all continuous estimation problems.

A third application is clustering. Typically, a population must be sorted and grouped based on common characteristics. One can identify market segments for consumers, or students with similar strengths and challenges, or words that belong to similar semantic groups. In its broadest sense, clustering also includes recommender systems, prescribing the next product to offer to a customer.

Another important application of Machine Learning is skill acquisition. DeepMind first rose to public fame by having an AI play Atari arcade games. The AI quickly reached superhuman performance and in passing devised winning strategies that the developers had no clue of. A robot can learn to fold and unfold laundry, to catch a ball or to climb a wall. A car might improve its navigation and trucks could optimize their delivery sequence. Traffic lights in a neighborhood may be coordinated to avoid congestion.

Segmentation of Machine Learning by learning paradigm:

There are different paradigms to train algorithms. In Supervised Learning, data is labeled by humans and fed into the algorithm. Recent applications of supervised learning on images, video or sound are innumerable: from detecting heart diseases on eye scans to sorting ripe vegetables, from doing presence at school to adjusting course content based on engagement, from analyzing mood on a tourist site to enforcing law and order, from detecting sentiment in customers’ voices at a call center to relieving patients in psychological distress. With Supervised learning, all data is labeled and the algorithms learn to predict the output from the input data. (Classification and Regression)

Unsupervised Learning uses data that the system is typically tasked with clustering or reducing dimensionally. The process is said to be unsupervised because it determines by itself the proper categories (if told how many there should be). For instance, if retail data is fed into an unsupervised system required to extract four customer segments, it will be able to do so without additional help. In multi-dimensional problems it can also figure out which dimensions are the most meaningful, ie would not distort the information too much if the other dimensions were discarded. With Unsupervised learning, all data is unlabeled and the algorithms learn to inherent structure from the input data. (Clustering and Association)

In Reinforcement Learning, the system is simply given an objective but not shown how to achieve it. After completing the task it receives feedback (the “reward”): was the objective achieved, or not? By repeating the process thousands or millions of times, the system develops strategies to execute the task correctly. (Skill acquisition)