Understanding AI: What AI is and how it works?

Before we can describe how AI works, it is important to first define the concept properly. AI is often referred to as a General Purpose Technology (GPT). GPTs are technologies that can drastically change an entire economy. They have the potential to disrupt societies through their impact on existing economic and social structures. Think of electricity, automation and the internet. Yet AI is not actually technology. It is the outcome, or goal, of an ecosystem of different technologies. A broad distinction can be made between three overarching technologies that can make AI possible:

> Whole brain emulation (WBE), also known as mind upload. WBE is the hypothetical process of scanning the entire human brain, including long-term memory, and then copying it to a computer. Theoretically, a computer could then run a simulation model of the brain's information processing, essentially operating in the same way as the original brain and experiencing consciousness. The Human Brain Project (HBP), one of the two largest scientific projects ever funded by the European Union, has been working on this since 2013 with 500 scientists and more than 100 universities, hospitals and research centers.

> Brain-computer interfaces (BCI) is a technology in which brain signals can be measured and digitized. A computer can then classify these and convert them into actions. In this way it is possible, for example, to control robotic prostheses.

> Machine Learning (ML) is about creating algorithms that can learn from data. People often say that a computer only does what it is programmed to do, but that is not always the case anymore. Machine learning is about a revolution in which people no longer program (if this, then that), but in which machines themselves derive rules from data. A machine learning algorithm is able to independently extract patterns from data, build models and make predictions about various things without pre-programmed rules.

When people talk about AI, in most cases they are talking about machine learning applications. This is without a doubt the most dominant association with AI. Research by the World Intellectual Property Organization (WIPO) in 2019 also shows that machine learning is the most dominant AI technology included in patent applications. Within this foresight, we therefore mainly focus on machine learning. This may seem like a limitation, but it certainly isn't. The applications of this field are almost infinitely wide and so complex that even the AI ​​experts disagree.

A dominant concept in this field is currently Deep Learning (DL). Deep learning is a machine learning method that uses different layered artificial neural networks. These algorithms are inspired by the way neurons work in our brain. Deep learning algorithms can handle large amounts of unstructured data and are widely used in applications in the field of image recognition, among other things. Deep learning has been the main driver of AI's success in recent decades and is largely responsible for the recent "hype" surrounding AI.

Cousins of Artificial Intelligence’ – Source: Towards Data Science (2018)

The AI ​​technology therefore does not exist. AI consists of a combination of technologies. You can roughly say that at least three components are needed, namely:

• Data: development is big data
• Algorithms: development is deep learning
• Computing power: development is Graphics Processing Unit (GPU)

The development of these components is closely related. Due to the increased availability of data, deep learning algorithms (which were already devised in the 1960s) have become increasingly relevant. This requires processors that can perform not only sequential, but also parallel calculations (GPUs).

The Definition of AI

The overarching nature of AI means that the concept is not easy to define. This is also reflected in the opinion of the European Economic and Social Committee (EESC). In the report from 2017, rapporteur Catelijne Muller states that there is no unambiguously accepted, well-defined definition of AI and that it is a container term for a large number of (sub)domains. 

“AI are intelligent systems that can perform tasks independently in complex environments and improve their own performance by learning from experiences.”

An important condition is therefore that the system must be able to learn independently. It is not a static model; it can adapt to practice. In other words, it should get better and better the more it is used. However, not everyone seems to be able or willing to associate with the concept of AI.

 » AI is what AI people do. «

–– Leendert van Maanen, University of Amsterdam

Some scientists and developers identify very strongly with the term AI, while others definitely don't want to. However, they basically do the same job. This says a lot about the reputation of the term 'AI'. AI is a bit of an unfortunate concept. The term is on the one hand stuck everywhere (while sometimes it is a 'simple' statistic) and sometimes it is just removed (for example with navigation systems). This is also known as the AI effect. We attribute properties to it and judge it before it has even been able to prove itself. At the same time, we decrease the properties if they are proven. 

» AI is whatever hasn't been done yet. «

–– Larry Tesler, 1970

For years we have thought that playing chess requires a high level, and above all human, form of intelligence. The insight and intellect that this requires could not simply be matched by a computer. Until in 1997 IBM's chess computer Deep Blue defeated the world chess champion Garry Kasparov. From this point on, we don't think chess is intelligent anymore. It is 'just' a formalization of possible moves and a computer with a lot of computing power can easily calculate this. It is not inhumane to assign ourselves a special role in the universe. Degrading AI successes helps us maintain our unique position. Once the mystery is solved, we prefer to call it automation, rather than intelligence.

It is therefore interesting to look at the different approaches to AI. Stuart Russel and Peter Norvig, in their 2010 standard work Artificial Intelligence: a modern approach, distinguish between four different approaches:

> Thinking Humanly

The automation of thought processes that we associate with human thinking, such as learning, problem solving, and decision making.

> Acting Humanly

Creating machines that perform actions that require intelligence when performed by humans.

> Thinking Rationally

Developing calculations and models that enable reasoning.

> Acting Rationally

Designing artifacts that exhibit intelligent behavior.

These different approaches underline the complexity of AI. The concepts of computers, machines and robots often overlap. Yet in most definitions of AI it is more about the 'mind' than about the 'body'. Robotics and automation are not necessarily AI. Therefore, in this foresight, a clear distinction is made between AI and robotization and we therefore mainly focus on the software, unless stated otherwise. 

The Classification of AI

Despite the fact that there is no agreement on the definition of AI, the classification of AI is widely supported. A distinction can be made between three different levels of AI development:

> Artificial Narrow Intelligence (ANI)

ANI is a form of AI that is very good at doing specific tasks. This includes playing chess, making recommendations and giving quantifiable predictions. Current applications in the field of image and speech recognition are also narrow. In fact, all the applications that we currently know of AI are specialized. We are not yet able to develop AI that can perform a task that transcends its own domain. For example, when you ask a navigation system to tie your shoelaces, it will not act.

> Artificial General Intelligence (AGI)

AGI is also known as human-level AI. This form of intelligence should be able to perform all intellectual tasks that a human being can perform. In addition to recognizing patterns and solving problems, this also means flexibly adapting to new environments. It's the dot on the horizon. But the further we delve into it, the more we realize that we don't have a good idea of ​​exactly how human intelligence works. Experts do not agree with each other. Some say general AI is just around the corner, while others say it will be nearly 200 years away.

> Artificial Super Intelligence (ASI)

ASI can be achieved when AI exceeds the capabilities of the human brain in all possible domains. So also in the field of science, creativity and social behavior. Even though this still seems very far away – we have not reached AGI yet – there are scientists who argue that the step from AGI to ASI is relatively small. For example, a Seed AI, a form of AGI that can improve itself by rewriting its own codes, could cause an intelligence explosion. Oxford professor Nick Bostrom describes in the book Superintelligence; Paths, Dangers, Strategies how we can survive this explosion of intelligence.

How AI Works

The idea of ​​intelligent machines dates back to the 1950s. In 1950 Alan Turing, key figure in breaking the Germans' Enigma code in World War II and founder of the computer, wrote an article on Computing Machinery and Intelligence. An important insight from this, which was later also adopted in AI, is that our brain and our mind relate to each other in the same way as a computer to a computer program. This implies that our brain is an information processing system and that thinking is a form of computation, or a mathematical calculation. His point of departure was that you can calculate anything with a computer and therefore also display intelligent behavior. For this, in the same article he introduced the so-called Imitation Game, later known as the Turing Test. Instead of asking whether machines can think, he asks whether machines can imitate intelligent behavior. A computer would pass this test if people don't know whether they are communicating with a human or a computer. 

» Thinking is calculating. «

–– Frank van Harmelen, VU

AI is basically 'just' mathematics. It is a very advanced form of mathematics, but it remains mathematics. Above all, it is a means to an optimization goal. This is all about decisions. For example with the self-driving car: the goal is to get from A to B safely. To achieve this, decisions have to be made continuously. That's basically how it works in an organization. Often the goal is profit optimization. Sometimes the solution is deep learning, other times an old-fashioned regression analysis will suffice. For example, if you want to optimize your sales as an ice cream parlor, it is useful to compare your sales to the weather. You will no doubt come to the conclusion that the weather and ice cream sales are related. You can therefore also predict that more ice will be sold when the weather is nice. Is this AI? No, but it is an algorithm if you formalize it.

An algorithm is therefore a mathematical formula. It is a finite set of instructions leading from a given initial state to a predetermined goal. For example, Facebook benefits from its advertisers' posts being liked by users. So if you often liked football posts, the algorithm will increasingly show similar content on your timeline. A neural network is also a mathematical formula and therefore also an algorithm. The algorithm of a neural network contains a component through which it itself learns patterns from data.

AI has evolved in several directions over the past decades: 

Good Old Fashioned Artificial Intelligence (GOFAI)

In the first phase of AI, from 1957 to the late 1990s, the emphasis was mainly on symbolic representations.

The premise here is that our brain is a symbol manipulator and follows step-by-step processes to parse representations of the world with calculations. This approach is therefore also called Symbolic AI. Formal logic, or if-then rules are used here. This means that a computer is able to make valid deductions from knowledge the computer already possesses. So, for example, if the computer knows that A is equal to B and also that B is equal to C, then the computer must be able to deduce that A is equal to C. It was therefore also called a knowledge-based system or expert system, where the system knowledge of human experts to solve a specific problem. This form of AI was mainly focused on high level cognition, such as reasoning and problem solving. Such tasks are very difficult for humans and people were therefore very impressed by this form of AI. Chess is a perfect example of this. Chess takes place in a boxed world, you don't need general knowledge to understand the rules of chess and you don't need to recognize the chess pieces. An abstract representation and brute computing power are sufficient. From that point of view, it was no shock that Garry Kasparov was beaten by IBM's Deep Blue in 1997.

Machine Learning (ML)

When NASA wanted to explore Mars with a computer-controlled vehicle in the late 1990s, they found that it could not be controlled remotely. The great distance caused delays in communication. This created the need for an autonomous vehicle that could independently avoid obstacles and ravines. NASA therefore decided to call in AI scientists. However, they had the greatest possible difficulty in solving this problem using Symbolic AI. When there are many different situations that an AI application can face, it is almost impossible to build the rules for all those different situations into the system. This problem is also known as the explosion of the state action space. Therefore, from the 1990s on, other approaches were looked at. An important scientist in this development is Rodney Brooks. In 1990 he wrote the article Elephants Don't Play Chess and introduced the subsumption architecture. In this case, the representation was not in the robots, but the environment itself was the representation in which they learned to maneuver. Instead of manually programming thousands of rules, machines learned to derive rules themselves from large amounts of data.

This created a paradigm shift from Symbolic to Sub symbolic AI. Sub symbolic means that there are no longer rules that can be expressed in words, but that the system has learned from many different examples what the rules should be. In this new phase of AI, also called Nouvelle AI, the focus was on low level cognition, such as perception and recognizing patterns, in contrast to Symbolic AI. This requires large amounts of data and training sets. Within machine learning, there are different ways for AI systems to learn.

A distinction is made between three types:

1: Supervised Learning is about carefully compiling and pre-labeling training data. The algorithm then learns to classify data on the basis of labeled examples. For example, if you want to teach an algorithm to recognize a car, you can feed it with thousands of examples of cars labeled 'car'. At the same time, you should label a large number of images without cars with 'no car'. It helps then to use related images, such as a tractor and a bus. When the algorithm has been trained, it can indicate whether or not it is a car on the basis of new images. The same principle applies to texts. For example, spam filters and translation machines work on the basis of supervised learning.

2: Reinforcement Learning is about learning through trial and error. The examples are not pre-labeled, but the algorithm receives feedback afterwards based on the results. When the algorithm proposes the correct solution, it receives a 'reward'. The advantage of reinforcement learning is that the system can control processes. For example, it is often used in training game systems, such as chess computers. Game situations can be easily simulated in a computer and played at high speed. The software of self-driving cars can also be trained in this way.

3: Unsupervised Learning is about algorithms that learn to search for clusters themselves in large amounts of unlabeled data. The algorithm then independently looks for similarities in the data and tries to recognize patterns in this way. In advance, only a measure is indicated with which the distance between data samples can be measured. The algorithm structures the data in such a way that data samples with a small distance from each other end up in a cluster. So cars with cars and cats with cats. Unsupervised learning can also cluster groups of people with similar preferences and characteristics based on user data. Many recommendation systems, such as the recommendations at Spotify and Netflix, work in this way.

Deep Learning (DL)

Deep learning is a much-discussed method within machine learning. Deep learning is a modeling method in which neural networks are central.

These artificial neural networks are originally based on the human brain, in which neurons are layered together. Artificial neural networks are composed of stacked layers of "nodes" that are interconnected.

Before a neural network is fed with data, random values between 0 and 1 are assigned to the different nodes. These values represent the intensity of the connections. When the network is fed with data, the intensity of the connections is fine-tuned by the network. Deep learning works just like neurons in the human brain that are activated by specific signals and transmit information. This allows the network to classify or cluster images of dogs, for example. 

Visualization of an artificial neural network

A distinction is made between input layers, intermediate hidden layers and output layers. At the input, the image is broken down into pixels. Different layers are able to recognize different properties. At one level, the system learns to distinguish between all kinds of lines, stripes and shapes. At another level, the system learns to recognize parts, such as ears and eyes. This allows the system to distinguish different images from each other. Deep learning is a counterpart to logical reasoning and tends more towards a form of intuitive learning. The disadvantage of this is that the considerations made within the system are much less easy to trace. In addition, a deep learning network can create representations of the data that are no longer recognizable to humans.

Image recognition applications often use a so-called Convolutional Neural Network (CNN). A CNN works like a filter that moves over an image looking for the presence of certain properties. Although the architecture for this is pre-designed, the filter values ​​are learned independently and automatically. Such neural networks are also called Feedforward Neural Networks (FNN). Although the different nodes are connected to each other, they do not form a cycle. They therefore only communicate with each other in the forward direction and do not store anything. Recurrent Neural Networks (RNN), on the other hand, are able to use their internal state, or memory, to process sequences of input. For example, the Long Short-Term Memory (LSTM) network uses so-called feedback connections, which makes it possible to create a loop. As a result, it is able to process sequences of data such as speech and video in addition to single data points, such as images.

A breakthrough in deep learning that requires less data is the Generative Adversarial Network (GAN). A GAN can be seen as two deep learning networks that are in competition with each other. It is a form of semi-supervised learning. One of the two networks is called the generator. It attempts to generate new data points that mimic the training data. The other network, the discriminator, retrieves this new data and sees whether it is part of the training data or whether it is fake. This creates a positive feedback loop, as the discriminator gets better at distinguishing original and fake data and the generator gets better at creating convincing fakes.

Within deep learning we see more and more combinations of learning emerging. For example, Deep Reinforcement Learning. For example, drones are trained to stay on a path. It is a combination of visual navigation and reinforcement. Even in the field of old-fashioned games, this results in impressive breakthroughs. Google's DeepMind created AlphaZero, a program that taught itself how to play Chess, Sogi and Go just by playing against itself. The algorithm was not fed with thousands of sample jars, but only the rules were explained. Within a training period of one day, it managed to beat the best computer programs.

The principle of neural networks is anything but new. The principle was first described in the 1960s. However, due to the limited computing power available, successes were not forthcoming and it has been in the refrigerator for a long time.

Functional Applications of AI

AI is basically a toolbox with different types of tools. These tools can be applied in various fields such as robotics, speech recognition and recommendation systems. In some cases, application areas are described as technologies in their own right, as there is overlap between computer technology, cognitive science and psychology, among others. For this foresight, however, it has been decided to refer to these application areas as part of machine learning, since machine learning methods are the main drivers of these techniques. By definition, AI cannot do without insights from other adjacent fields and transcends domains. For the overview, a selection has been made of the most discussed application areas of AI:

> Natural Language Processing (NLP) is about a system's ability to understand human language. An important aspect here is semantics, or the meaning of symbols that form the building blocks of natural languages. Applications can be found in chatbots, translation engines, spam filters, automatically generated summaries and search engines.

> Speech Processing is about systems that can recognize and process speech signals. This enables voice searches. Think of the different voice assistants such as Siri and Alexa. A distinction can be made between speech recognition and voice recognition. Voice recognition allows people to be identified by their voice. It is also capable of creating artificial voices.

> Computer Vision is about systems that are able to extract information from images and thus recognize them. The goal is that these systems can ultimately understand images and interpret their content. Image recognition is already being used in various domains, such as medical science (including screening lung photos for abnormalities), multimedia (including visual search engines) and defense (including surveillance). In addition, it is possible to artificially adjust and even create images. These are also called DeepFakes.

> Affective Computing is about systems that can detect and recognize emotions. This application is on the rise, especially in the marketing world. Many purchases are emotion-driven. Our face and attitude often betray our emotions. It can therefore also be applied in police work. For example, a robber could be recognized on the street before he or she commits the crime.

Learn more about AI here.