Once started as a scientific trial, and now a hot item in technological and arithmetic circles: machine learning. This branch of sport arose from the endless search for a form of artificial intelligence.
In the meantime, science facts have arrived at a point where science fiction is slowly coming into contact. Not yet at the point where independently operating androids, à la C-3PO from Star Wars, dream about electric sheep or have to deal with an existential crisis. However, we are getting ever closer to what was once sketched as the future image of artificial intelligence.
WHAT IS MACHINE LEARNING?
A commonly used, formal definition of machine learning is a technique where ” a computer program could learn from event E, relative to similar tasks T and performance measure P, as its performance on tasks in T, as measured by P, enhanced by experience E. ”
Machine learning includes, in short, computer algorithms that are used to learn data and input autonomously, so without supervision. This means that computers do not have to be programmed themselves, but can change and improve their algorithms independently.
Today, machine learning algorithms are used to, among other things,
- allow computers to communicate with people,
- enable self-driving cars,
- write and publish sports competitions’ reports and statistics,
- check emails for spam messages,
- timely diagnose sick patients,
- and to identify and locate potential terrorists.
Virtually every industry will be confronted with machine learning and will immediately discover the positive consequences.
THE ROLE OF MACHINE LEARNING
Already at the start of the artificial intelligence hype, several researchers realized the importance of machines that can independently learn from data – which so far has proved to be one of the cornerstones of AI. Research into neural networks, statistics and probability calculation as part of intelligence has so far led to enormous advances in the field.
Over time, the paths of machine learning and artificial intelligence began to separate slowly. The increasing focus on an approach based on logic and knowledge-based model pushed the probability model further and further into the background: IOT Machine Learning and thus the role of machine learning. Probability calculation, after all, had to contend with a lack of confidence. Practical problems such as the collection of the correct data and the proper processing of these led to a substantial decrease in the use of statistical models.
At that time, machine learning moved away from artificial intelligence, and more towards programming distracting logic systems and systems for probability and statistics: ideal for retrieving information and pattern recognition. ‘Intelligent’ algorithms can learn from enormous amounts of input – and convert this into sound output, whether or not as a result of a new specific data set that must be connected to a specified output.
Such algorithms can thus ultimately make predictions or make decisions based on the collected data and the analysis thereof. As the examples mentioned earlier of implementations of machine learning already demonstrate, it is therefore at its most valuable for applications for which manual design and programming of explicit algorithms are extremely difficult or even impossible.
FROM THEORY TO REALITY
Arthur Samuel, an IBM employee, is the one who uses the term machine learning I introduced: this was in 1959. As an employee of his company, he was seen as an authority and pioneer in the field of gaming and artificial intelligence. He used a chess game for this, which became increasingly smarter as it played more. It remembered winning moves and used them in its games.
This was the beginning of the elaboration of the concept of machine learning in general, and the chess computer in particular: it is therefore regularly regarded as the grandfather of Deep Blue, the IBM computer that defeated chess champion, Garri Kasparov, in 1997.
Many innings followed after Samuel ovations that all went on to find algorithms that teach themselves, based on the collected data, and based on which they are capable of independent decision-making.
A highlight of this took place in 1958, the year in which Frank Rosenblatt designed the Perceptron. This was the first artificial neural network a copy of the human neural network – which formed the basis for more human intelligence in computers. An analysis was carried out on the basis of stimuli (input) and then converted into a reaction (output).
Also worth mentioning is the project by students of Stanford University in 1979, which led to the development of the so-called.’
Stanford Cart ‘led: a moving robot that was able to move independently through a room and avoid obstacles along the way. A distant ancestor of the self-driving car and robot-vacuum cleaners.
TYPES OF ‘LEARNING’ WITHIN MACHINE LEARNING
A computer can learn in different ways. Such smart algorithms can be classified into several broader categories based on the way they learn. This includes the following learning methods:
- Controlled learning, in which the algorithm gets examples of standard input and the corresponding output. Based on these examples, the system learns how specific characteristics of the input determine what the output will be. After the initial learning phase, such an algorithm can independently convert new input into the correct output, or exercises are performed in which certain input elements have to be divided into groups.
- Independent learning, with no examples of current input and the desired output, but the system itself learns, based on the structure of the input – for example, by subdividing input into similar groups.
- Semi-controlled learning contains both elements of controlled and uncontrolled learning. After an incomplete ‘learning session’ via the controlled method, the system further learns through independent learning.
- Supported learning teaches behaviour to itself, based on its relationship to the rest of the world and achieved successes. The system thus learns by doing, for example, by driving in a vehicle or playing a specific game.
- Transduction learning is the least used of all methods, and can only be trained for a limited set of cases (based on a budget), whereby a choice must be made independently between cases in which training is required.
- Deep learning uses the input as the basis for understanding ‘the world’ and developing concepts. For example, for recognizing shapes, or an animal: does it have straight sides? How much? Does it have eyes, ears, what shape? Based on such input, deep learning is used to recognize concepts and to reflect – after recording new data on new concepts.
APPLICATIONS OF MACHINE LEARNING
Some of the most used applications of machine learning technology are for tasks that are almost always identical in execution, i.e. repetitive work, whereby only a limited number of outcomes are possible. Based on input variables, a system can continuously develop itself to perform its task better and deliver reliable output.
Several specific fields in which machine learning is already used include:
- Bioinformatics, with which protein functions can be predicted, among other things.
- Self- driving vehicles, which can go entirely autonomously from point A to B, taking into account all variables en route – such as traffic, circumstances, obstacles, traffic rules, etc.
- Natural language processing, so that for example, the function and meaning of words in a sentence can be determined and more effective translation and writing programs can be designed.
- Audiovisual data processing, whereby visual representations and audio fragments – for example, captchas – can be analyzed, and after recognition has taken place, an output can be supplied that matches the audiovisual data.
- Face and voice recognition, which, like the example mentioned above, store specific characteristics based on visual or audio input and match this to matching output. A face recognition system is an excellent example of this, where input from many surveillance cameras can be read out to find similar matches with a predefined profile.
- Email spam filters, where a system teaches itself which email messages are potentially unwanted and spam.
- Sports analysis, in which the system takes over the role of a reporter or point counter and provides output in the form of a written and published match report based on input – in the form of measurable events on the sports field.
In theory, therefore, any industry that benefits from simplifying repetitive tasks that are difficult to automate could use machine learning. An algorithm can classify the often subjective and multi-variable input into usable output, which would save a considerable amount of time.
Something that is a precursor to accurate artificial intelligence, although due to its stubborn focus on the pattern or probability elements, it also deviates somewhat; nevertheless, such an application – in its field – is rightfully called intelligent. It is sometimes stated that a form of machine learning can be classified as a ‘first generation’ robot: it is not entirely unjustified, as the tasks it can perform.