What is Artificial Intelligence ? How does Deep Learning work? An overview

Which technologies are hidden behind the term artificial intelligence or AI? What can machines do better than humans? And do we need to worry?

Artificial intelligence has created self-learning systems. Meanwhile, humans must learn to use artificial intelligence correctly. A child interacts with a robot. Hong Kong Science Museum, 2021.

What is artificial intelligence?

The dream of making machines intelligent is hundreds of years old. With the development of computers, new hopes were placed in its fulfillment. Computers were programmed with expert knowledge that they could call up and apply with increasing speed. This is how early forms of chess computers came about.

Today, computer programs also take on more complex tasks. This includes things that were once thought only for intelligent beings, such as translating, spotting tumors on X-rays, or driving a car. Such programs are commonly referred to as artificial intelligence or AI.

The expert group of the EU Commission provides a narrower definition of artificial intelligence: It analyzes its environment and takes actions to achieve a specific goal - whereby it acts with a certain autonomy. Accordingly, one always speaks of artificial intelligence when not every action of the computer program has been previously defined and programmed.

What is machine learning?

Machine learning is a special form of artificial intelligence: in this case, programmers specify the strategies according to which data is analyzed, but they do not specify every detail as in conventional programs. Instead, within a given framework, the algorithm independently improves its approach to data analysis.

One application of machine learning recognizes faces in images. Earlier computer programs, which could not learn themselves, were programmed with rules by which a face could be identified. In other words, what constitutes a face was “ explained ”: for example the shape, the color, and the position of the eyes or mouth. However, since images come in so many forms, such programs were not very effective.

With machine learning, the algorithm processes sample images to learn what makes a face. There are several methods: In so-called supervised learning, training images are " labeled ", so the algorithm learns from the start which images are in the " face " category. By comparing it with images of other objects, he then learns what distinguishes faces.

However, there are also learning methods that completely dispense with this categorical information. Then the algorithm identifies the categories of the training images itself and can come up with solutions that humans would not have thought of.

A self-learning system can be based on different types of algorithms. One of them is the so-called artificial neural network.

What are artificial neural networks?

Artificial neural networks are systems of algorithms that are based on the way the brain works. So it's not a physical replica of natural neural networks, but a replica of how the brain works and learns.

The human brain consists of a network of different nerve cells that are connected to each other by a large number of cell processes. Learning in this network structure means that certain connections become faster and others slower.

Similar to the brain, artificial neural networks can weigh information in complex structures and continuously adapt this weighting to new requirements. This is possible thanks to mathematical optimization functions with a huge number of variables. The learning method of these networks is called deep learning.

What is Deep Learning?

Deep learning is a machine learning method that uses artificial neural networks. For better understanding, these can be imagined as a network of nodes and connections. The nodes of such a network are organized in layers, as shown in the figure below. As you learn, the links between nodes become stronger or weaker. The strength of a connection is expressed by a value between 0 and 1.

Artificial neural networks are used in very different areas of artificial intelligence. Here we explain how it works using image recognition.

The first layer of an image recognition neural network contains roughly the color of each pixel in an image. The following layers link the information to form more complex relationships. The last layer contains the result, in this case roughly what is shown in the picture. With practice, the system learns by itself which information is important. It analyzes tons of images, adjusting its links and nodes as you go.

However, it is largely unclear what exactly happens in the neural network. It's to some extent a black box that you can't look directly into. As a result, it can happen that inappropriate criteria are used to differentiate between image categories. For example, an algorithm " recognized " the difference between dogs and wolves in the snow that was visible in the background of the image, because in the training data wolves were almost always depicted in a snowy landscape. Such a system can perform well on training data, but fail when used in the real world.

What can the largest neural network in the world do?

The largest neural network ever trained is called GPT-3. It has learned to write texts that at least partly appear as if they came from a human being. It is built on an algorithm that predicts what the next word is when given text to it. The predecessor, GPT-2, can be tried out here, for example. These networks were trained by reading practically the entire English-speaking Internet.

The GPT-3 network has 175 billion parameters, i.e. values ​​to be adjusted during training. This is an incredibly complex math problem. Therefore, such large neural networks with many layers are criticized for their high energy consumption. However, once trained, they can also be adapted for other problems, in this case summarizing texts.

What can AI do better than humans?

Computerized algorithms nowadays process a large amount of data at the same time. As a result, they surpass the computing power of a human brain many times over. For example, it is impossible for humans to defeat a chess computer that can calculate millions of chess moves and their consequences in a fraction of a second.

Today, AI can detect tumors on X-rays, suggest the next suitable song to music listeners, or advance scientific research. For example, a neural network made it possible for the first time to predict how a protein will fold based on a given sequence of amino acids.

In which areas are humans superior to AI?

Overall, the brain processes information more slowly than a computer. But thanks to decades of the learning experience, the human brain can still solve problems in a short time. It is superior to the computer, especially when it comes to solving a completely new type of problem.

How the brain proceeds in detail is the subject of neuroscientific research. Cognitive researchers speak of so-called heuristics. These are rules of thumb that people use to solve problems. The characteristic thing about it is that instead of processing all available information, the human brain uses only a few information-rich clues. In this way, it achieves the optimal balance between speed and perfection – a solution that is both fast enough and good enough.

Simple heuristics can be essential for survival in everyday life. For example, our brain has learned that a moving object can mean danger. This allows us to evade an approaching car in a flash without first processing irrelevant information such as the color of the car. Heuristics allow us to deal effectively with the limited computing power of our brains.

Computer systems are still a long way from being as efficient in handling the wealth of information that makes up the real world.

Will there be human-like machines soon?

A system that not only takes on specialized tasks but can do a wide range of things like humans, is called “ general ” artificial intelligence or “ strong ” artificial intelligence. From computer science to philosophy - there is a debate from many perspectives as to whether such systems are possible in principle.

One wonders, for example, whether one can only speak of genuine intelligence when consciousness is present. Then machines might never become intelligent. Linked to this is the question of what constitutes human consciousness in the first place. Does this include the ability to think about yourself? Or does human-like intelligence also include a will of one's own, an individual motivation to act?

Many experts are convinced that machines will achieve human-like abilities. Forecasts like this have been around for a long time, and up to now, they have regularly been pushed back. For now, one of the things they're basing it on is the rate at which chips have gotten faster and cheaper in recent years. If this development continues, so the argument goes, the computing power of the human brain could be digitally reproduced in thirty to forty years, i.e. around 2050 or 2060.

But there are arguments that speak against the fact that there will be human-like machines soon - or at all. For example, that so far progress has only ever been made in some areas of AI, but we are still a long way from flexible adaptability like that of the brain. Or that learning also has motivational aspects that cannot be reproduced. In addition, there is the above-mentioned question of whether machine consciousness is possible.

How is AI already influencing our everyday lives?

As is often the case with AI, the law lags behind technical development. Because AI systems can develop further during use and because it is often not clear from the outside how exactly they come to certain conclusions, it must be clarified who is responsible for their decisions.