by Nikos Giannaros*

The history of Artificial Intelligence as something achievable coincides with the history of computer evolution. The first theories for the development of systems with Artificial Intelligence capabilities came out back in the 1950s. Only man possesses such capabilities and could be able to teach a computer system to take autonomous decision and act as a thoughtful being.

The initially theoretical approach of Artificial Intelligence was directed to the creation of the necessary tools for a computer to act like the human brain.

These tools were based on logic and semantics, which constitute the prominent tools with which the human brain thinks and acts depending on the stimulus it receives. In this context, there have been efforts to standardize logical rules and correlations, which led to the creation of tools such as Logical Programming.

The basis of Logical Programming is the provision of basic entities to a computer system, so that with the application of a procedure of creation of logical conclusions, the system will be able to take decisions, imitating the procedure followed by the human brain.

This approach is based on the “up to down” logic. The system is supplied with the total of the requisite knowledge and the endeavour focuses on the way with which it will manage and combine properly this knowledge.

It soon became obvious that this kind of approach can offer limited practical solutions, because are not harmonized with the fundamental functional principles of a computer. The computer in its core is a machine, which can process mathematical calculations with great speed. There was a need for an approach which would work the other way round; the “down to up” approach.

This approach focuses on the endeavour to modelize a problem of -apparently- logic to a pure mathematical problem. A mathematical problem can be solved by using certain procedures with clearly determined steps, on which the computer is highly efficient. The outcome is also easy to classify, to categorize and be interpreted by the computer.

This finding led to the evolution of Machine Learning since the 1980s. This approach aims at modelizing through math certain human actions, so that a computer can conclude to the same result.

The most notable of these problems include the recognition of voice and image, processing of a human language into written form, robotics, forecasting, etc. Machine Learning uses the power of computers, their incomparable calculating capability, to make good use of big volumes of data to learn successfully to take decisions such as the ones taken by humans.

The models used are based on the way of functioning of the human organism (neuronic networks, genetic algorithms), of other living organisms (particle swarm optimisation, ant colony optimisation, bees algorithm), on applications of probability theory and of other field of Mathematics.

The various approaches which result to tangible algorithms, which can be carried out be a computer, that does not need any special structure or design, was revolutionary for Machine Learning.

Machine Learning consists of 3 main fields:

– Supervised Learning;

– Unsupervised Learning and

– Reinforcement Learning, which constitutes a combination of the other two.

The notion of Supervised Learning is the simplest of the three; the computer is provided with an appropriate training data set to get trained on the work which we want to assign to it.

The most common example for Supervised Learning is the provision of photos of men and women to a computer; on the photos there is a tag clarifying whether the photo depicts a man or a woman. A successful training will result in the computer being able to recognize the sex of the person depicted in any photo, since it will have been trained in classifying the data it receives.

In order to achieve this, the computer is provided with many features for every sex. Therefore, it becomes able to decide upon the sex of the person depicted in any photo with remarkable precision, based on the features appearing on each photo. This procedure requires that the determination and standardization of the features has been made by a human, specialized on the topic, before the training begins.

The feature extraction and selection constitutes a keystone for the successful application of Supervised Learning and has ended up being a whole, autonomous field.

Supervised Learning has long got away from the experimental and theoretical sphere and is used in commercial applications of voice, face and handwriting recognition, as well as in more specialized applications depending on the applicable field, such as customer segmentation, recommender systems/collaborative filtering, stock market prediction, preventive maintenance, etc.

Furthermore, with the appearance of Deep Learning, a sub-field of Supervised Learning, the feature extraction and selection can now be carried out by the computer itself; this results in further diminishing human intervention in the procedure. In the example of photo classification, the computer could, after having processed the photos it used as a training set, decide that lip shape is a more vital characteristic for the classification than eye shape and therefore give more importance to this feature.

In Unsupervised Learning, the computer does not first come through a training stage with use of data with determined features, in which the intended result is known. On the contrary, it is self-trained based on unknown data, which it can process.

The computer’s potential to work without first having been provided with knowledge, makes Unsupervised Learning look like extremely exotic and part of science fiction.

The truth is that in this case also, the applicable models make use of the computer’s ability to compare things and decide the level of similarity between them.

The structural difference with Supervised Learning is that the computer does not “know” the qualitative interpretation of the produced result. In contradiction to the example of Supervised Learning, an algorithm of Unsupervised Learning, which can classify photos of people according to their sex, can do so with equal rates of success, but does not know which category belongs to “man” and which to “woman”.

Unsupervised Learning has also many commercial applications. These are often in fields like the ones of Supervised Learning. However, as the power of computer systems is rapidly evolving, the ability of a system to be self-trained will replace the onerous human preliminary work with one more automated procedure. A perfect example for Unsupervised Learning is the ability of a computer to classify vegetables and fruits according to their shape, their size and their colour, without having first been trained to do so.

Finally, Reinforcement Learning constitutes a combination of the other two methods and is mostly used in robotics. The robot tries to carry out the instructions with which it has been trained in the best possible way, but on the same time enjoys some freedom to decide to deviate from these instructions. This might result in an even better result than the one already achieved.

Applications of Reinforcement Learning, such as self-driving cars, are in the anteroom of commercial application. Moreover, the various robots, which are used as human assistants, apply successfully the model of Reinforcement Learning, since their behaviour evolves based on the various stimulus they receive from different people with who they discuss.

It is obvious that Machine Learning is already here. It constitutes a reality and part of our everyday lives and not a science fiction plot. Many projects which seem evident, such as the recognition of the plate of our car in the airport parking, the digital help assistants in the phone centre of banks, which recognize our voice instructions, the applications of recognition of the music track we listen to (such as Shazam), the advertisements and the suggested posts in social media (such as Facebook, which combines the posts we open with the history of our search engine to show us relevant advertisements), the security applications, which activate video recording when a man approaches the door of a shop out of working hours, constitute a reality thanks to Machine Learning.

The fact that it is impossible that a person carries out the same tasks in the same time, justifies the label of “Artificial Intelligence” that is given to such applications.

On the other hand, the fact that computers carry out these tasks mechanically, without the result having any meaningful result for them and without them deviating from the way they carry out a task if they continue receiving the same stimulus, means that Machine Learning is still far from becoming real Artificial Intelligence; there are many unknown steps until this happens.

Computers carry out the task they have been assigned to without judgement and without feelings, with no possibility to deviate from what they have been designed to do. They can carry out the the tasks assigned to them in an increasingly better fashion, but they are not able to “start a revolution” and change their way of operation. A future in which humans and machines will be equal is definitely far away. The reality of Machine Learning though, is here.

*Nikos Giannaros is an Electric and Computer Engineer. He specializes in Artificial Intelligence and Machine Learning. He is interested in the sociopolitical impingements of technology.