As promised from last week, this week we will learn about the three main categories of machine learning. Machine learning is categorized by the nature of the training “signal” and “response” available to the system. The three categories are supervised learning, unsupervised learning, and reinforcement learning.
Supervised learning means that the algorithm is trained using a set of data and the expected results from the sample data. This “training data” can consists of numeric values, strings labels, or a series of images. This data is utilized by the algorithm to later predict the correct response when given a new piece of data it has not seen before. This approach is similar to human learning under the supervision of a teacher. A great example is a student learning a new language. The student will be given a series of words in the new language and told the corresponding word in their known language. The more words they are trained with the easier it becomes to understand the new language. They may even develop the ability to guess the meanings of unknown words based on the surrounding words. Computers trained with supervised learning work in much the same way. The more training data provided to the algorithm the better the algorithm learns to predict the correct result. Supervised learning is all about giving enough sample data for the computer to form general rules about how to translate the input data to the correct expected end result. This category is the most widely used in data classification problems, like label pictures to identify plants, animals and objects.
Unsupervised learning is like giving the student a newspaper and a few movies in a foreign language and letting the student figure out the new language on their own from the context, scenery, and social interaction on the film. This is unsupervised learning in a nutshell. The computer algorithm is given a set of data and the goal is for the computer to find patterns within the data and draw it’s own conclusions. These algorithms are used to help us spot patterns that are not easily seen. One of the main uses of these algorithms are in marketing automation used to make shopping suggestions on websites, or display related advertising on search engines.
Reinforcement learning is similar to supervised learning except the algorithm is presented with unlabeled data, like the unsupervised version and given positive or negative feedback on the accuracy of the result. I am reminded of a scene in the movie Ghostbusters when Peter is trying to train a male student to access his psychic ability by applying a shock when the wrong card is guessed. Only in the movie he shocks the poor boy on every card. The main difference is that this allows the algorithm to learn in a prescriptive manner rather than learning only through description. This can be compared to learning things by trail and error. We can only learn to ride a bike, by riding a bike, and we crash many times while learning, this is reinforcement learning. Errors help the computer learn because they have a penalty, teaching that one course of action is better than another. The main place we see this category of machine learning is in game theory. Computers use this method to learn to play video games, they learn by experience good or bad outcomes based on their decisions. Google Deepmind uses this method to play and master classic Atari video games.
Next week we will cover the another classification system of machine learning algorithms based on the desired output of the algorithm. Until next week stay safe and learn something new.
Scott Hamilton is an Expert in Emerging Technologies at ATOS and can be reached with questions and comments via email to firstname.lastname@example.org or through his website at https://www.techshepherd.org.