The Generations of Artificial Intelligence
The development of artificial intelligence (AI) has taken place in three generations so far, in so-called “waves”. Classifying AI in such simple categories is a simplification, for which can be found numerous counter-examples. Nevertheless, three generations can be distinguished, which are characterized by the following features: 1) Handmade, 2) Statistical Learning, 3) Explanability and Generative Models.
First Generation: Handmade AIs
The first generation includes more or less handmade, intelligent systems. In the first wave, artificial intelligence lacks a self-taught knowledge base. Decision, optimization or search problems are therefore always solved by AI systems based on expert knowledge. The capacity for abstraction is therefore still very low. Because oft hat the approach is only suitable for certain classes of problems. For example, a first generation system can help a mail order company to save shipping costs: How do I have to divide the items of a large order into how many packages of what size so that the shipping costs are lowest when they are delivered?
Chess computers also belong to the second generation. The computer is programmed with the rules of the game and then independently calculates the best of all possible moves. In May 1997 the time had finally come and a chess computer called “Deep Blue” was able to defeat a chess world champion, at that time Kasparov, for the first time. However, the victory was only possible by using considerable computing power and special hardware.
Second Generation: Statistical Learning
For “Go”, a game known mainly from the Asian region, it took almost another 20 years until a similar success was achieved. In contrast to chess, Go is much more complex compared to the number of possible positions. This makes it impossible to identify the most optimal move just alone by looking at the game tree. Nevertheless, the program “AlphaGo” managed to defeat the world champion Lee Sedol in March 2016. It used techniques of deep learning, such as Deep Neural Networks.
Currently we are mainly talking about the second generation of artificial intelligence, which can be roughly described as statistical learning. While the technological foundations are much older, statistical and deep learning (in the sense of second generation AI) have been successful, especially since 2012. These include speech recognition systems from the field of machine translation, or everyday helpers such as Siri, Alexa and Google Assistant. Apart from the obvious and well-known achievements, AIs are already statistically better than a human being in less known areas of application, such as lip-reading. One also speaks of a sole statistical superiority because the problems to be solved are mostly uncertain decisions. With these, a person often cannot find a clear “right” or “wrong” answer at first go. Therefore it is important that fixed rules are given for statistical systems. Otherwise it would be much more difficult to make a statement and explain the behaviour of the system. For example, researchers have explained in a research report how neural networks can be misled by specifically generated patterns. For example, the machine recognizes a cheetah in an image that shows only an indefinable noise for a human being.
Outlook: Third Generation: Explainability and Generative Models
With the third and so far last generation of artificial intelligence, research is only just beginning. The researchers are pursuing the goal of creating a system that is not only capable of achieving a desired result, but also has the ability to explain the derivation of the result. In this context, the requirement for transparency is often discussed in AI systems. This means that the reasons for the decisions made are comprehensible and can be transferred into a form that is understandable for humans. With regard to the cheetah problem above, it would be conceivable not to teach the AI how a cheetah looks like by means of a vast number of examples, but to show how to paint a cheetah. The question: “Does this look like a cheetah?” would consequently change into: “If you were to paint a cheetah, could it come out anything like this picture here?”
As already mentioned, active research is being conducted on the third generation of artificial intelligence. There are still many detailed questions unanswered. To what extent the current generation will be as successful or even more successful than its predecessors remains to be seen. However, techniques such as generative models are promising.