Machine learning is a real achievement. Learning by experiencing behavior and remembering it is no different from learning by humans. In the article we will try to take a closer look at this very interesting fenomenon.
Machine learning and its definition. What is this?
Machine learning is one of the components of artificial intelligence (AI). Machine learning algorithms enable artificial intelligence to understand the collected data. On this basis, AI can make independent decisions without having to program them each time for a specific task.
Machine learning is divided into four types, which differ in the amount of marked data. Each of them is used in the analysis in other usage scenarios.
Machine learning methods
Supervised learning is about defining the correct answer based on the inputs and outputs. The first ones are all the elements that make up the file saved in the device’s memory. The second – only those that are the right answer.
An example of supervised learning is finding the right picture. Suppose we want to teach a machine to find photos of Bugatti Veyron cars. In the set of images of various car models, we mark a photo of this specific vehicle as the output data. The machine learning algorithm collects information about the similarities and differences between cars and on this basis can correctly search for other photos of the model.
Related: The Internet – what is it? All about the Internet, www and its possibilities
The advantage of this type of machine learning is the small amount of necessary data and the possibility of comparing the results with the previously marked correct answer. The disadvantage is that the learning outcomes may be closely related to the input information, which can make it difficult to interpret changing data.
Unattended learning is about finding patterns based on large amounts of untagged data. In this type of machine learning, phenomena are combined into sets. Based on their common characteristics, the algorithm learns to recognize similar things.
This type of machine learning can be compared to the way humans recognize objects and phenomena. Knowing what a cat looks like, for example, allows the human brain to classify similar-looking animals to the cat collection.
A middle ground between the two types above is semi-supervised learning. A small amount of marked data entered into the system is contrasted with the group of unclassified ones. The task of the machine algorithm is to analyze the marked information in order to find similarities with similar data in an unclassified set.
The last type is reinforcement learning. There is no single correct answer or group of tagged data in this model to find in untagged sets. The machine learns the rules on the basis of which one of the possible solutions may be correct. They can be compared to learning about the rules of the game. An example is poker. There is no way to predict all the possible plays of an opponent with whom you play cards. However, knowing the rules, the player (e.g. which pieces and their layouts are higher than which ones, how many tokens are worth placing in a given situation) can increase their skills and win thanks to them.
Augmentation is a kind of reward. For a human being, it is the satisfaction of winning by mastering the rules. The reward for the machine is numerical. He can “get” it by looking for one of the potentially correct answers.
History of Machine Learning
The beginnings of machine learning date back to the 1950s. The pioneer and creator of the concept is Arthur Samuel, a graduate and lecturer at MIT and an engineer working at IBM. He has investigated this issue since 1949. The fruit of his work was the creation and development of a program for training chess players in the years 1952-1962. He first used the term “machine learning” in 1959.
The task of the program he created was to develop new tactics based on the general rules of the game entered into the memory of the device. For this purpose, he chose checkers as a simple game, but not without strategy. The analogy with reinforcement learning is not accidental. It was impossible to record all combinations of pawns as there could be millions of them. A general analysis of the rules had to be enough to create new gameplay variants.
He created a tree with board positions available from a given state. The small amount of computer memory forced Samuel to use a trick. The alpha-beta reduction algorithm did not search through all possible solutions to a game of drafts until the end of the game. The engineer developed a scoring function based on the combination of pieces on the board at any given time. This function measured the chances of winning, taking into account various variables such as the number of pieces left on the board.
The program remembered each of the items and associated them with the resulting result. Knowing the rules and variants of moves, he was looking for new ways to win. Samuel perfected his work over the years until his program could take part in a duel with a human.
Machine learning was also researched in other academic centers. In 1965, the Dendral was created at Stanford. This system was used to automate the processes of analysis and recognition of previously unknown molecules of organic compounds. Dendral turned out to be a breakthrough. For the first time in history, discoveries were made not by a human, but by a computer.
Science was also revolutionized by AM (Automated Mathematician) and Eurisko by Doug Lenat. Programs written in Lisp were designed to find new mathematical laws.
The most famous example of the use of machine learning is a chess match between a machine and a human. In 1997, Grandmaster Garri Kasparov lost a game of chess to Deep Blue. Would he lose if he were allowed to analyze the game strategy of the IBM computer? Hard to say. The Russian accused the American company of dishonesty. Before the match, the machine had the opportunity to study Kasparov’s tactics. He himself could only make use of checking the operation of commonly available chess programs. Despite the grandmaster’s insistence, the rematch with Deep Blue was never achieved.
Nowadays, we deal with artificial intelligence every day, often not even being aware that it is so. Machine learning is widely used in the Internet, mobile technologies and business. The following examples may make many people realize that we come into contact with it almost constantly in our lives.
Application examples for machine learning
We can meet using the Internet. Machine learning mechanisms, remembering the phrases entered in the search engine, the history of browsed websites and the posted comments recommend websites, products and services that the Internet user may be potentially interested in. This is also how social networks such as Facebook and TikTok work.
Machine learning is also used by film and music websites. Based on the preferences of the Netflix user, Spotify and related sites recommend more songs, artists and movies.
We can meet this technology by talking to a voice assistant. Siri, Cortana, Bixby, Alexa, and the Google Assistant process human language to interact with humans in the most natural way possible. Machine learning is not limited to understanding spoken messages, but also written language and can answer human questions as a chatbot.
Machine learning is used in autonomous cars. Today, such vehicles are used, for example, by Google to create Street View in many countries. The vehicle constantly analyzes objects in its vicinity and traffic and adjusts it to the well-known provisions of the Highway Code.
The above-mentioned social networking sites also use machine learning to detect threats. This allows harmful content to be identified and blocked. This is crucial for moderators as people are unable to keep track of the millions of comments, videos and links shared by users.
These are not the only applications of machine learning, but only those with which we have the opportunity to meet on a daily basis. This technology is also used in areas such as medicine, logistics and others.
Deep learning and neural networks – the most important information
As mentioned at the beginning of the text, machine learning is one of the components of artificial intelligence. At the lower levels of this technology is deep learning, and below it is neural networks.
Deep learning algorithms trigger neural network layers to analyze large amounts of disordered data. This allows the device or program to perform actions more accurately and achieve the desired results.
Deep learning is used, for example, when searching for images and talking to a voice assistant. Here again reference can be made to the Bugatti Veyron search. Let’s assume that we want to find this car model in red. First, the system distinguishes photos of cars from others. Then he recognizes the Bugatti photographs, then the Veyron copies, and finally the red ones.
Voice assistants recognize the language first, at the deeper level of the word, and finally, whole sentences. By reading their meaning, they can communicate with a human being and answer questions or activate a desired function, e.g. turn on a light.
The deepest, elementary part of artificial intelligence is artificial neural networks (ANN). Their operation is similar to natural networks in a biological brain. After receiving numerical information, a single neuron transmits it to other neurons. The layers of the neural network are connected and function simultaneously. This allows for more accurate recognition of patterns and correlations between the collected data and gathering this knowledge for learning.
Machine learning is one of the most interesting inventions of the 20th century. It saves time for engineers who do not have to constantly program machines. It helps companies to reach customers more effectively and doctors to make more accurate diagnoses. It is also useful for easier searching for interesting content. Inextricably linked with the technology that surrounds us, machine learning has changed the face of science, business and entertainment.