What is Artificial Intelligence?
This article provides an overview of the variants of artificial intelligence.
Even after more than 60 years of research on the subject, there is no universally accepted definition of artificial intelligence, and due to various factors, artificial intelligence has been gaining tremendous attention in recent years. The common abbreviations AI and CI (for “Artificial Intelligence” and “Computer Intelligence”) are now ubiquitous and yet few know what it is all about. A university professor recently defined AI in his lecture on Artificial Intelligence this way: “AI is a subfield of computer science.” He didn’t let on much further. Often, and also by him, this is gladly justified by the fact that there is also no generally accepted definition of human intelligence. But this does not bring us further in the matter and therefore we begin with a short consideration:
A widely accepted rule is that intelligence is what intelligence tests measure. Well, but what is that?
Common tests are about autonomously analyzing questions (problems) and solving them. The acquired knowledge of mathematical rules, geometry and language takes a back seat. Knowledge has more to do with education and only to a basic extent with intelligence. Also, the tests are designed in such a way that they function largely independent of the performance of the sensory organs and the motor requirements for solving the tests are minimal. Thus we limit ourselves to tasks that the brain can solve completely autonomously. Whether imagination plays a role, we leave unconsidered at this point, in any case it is more than mere logic.If we now transfer this to AI, AI would be any machine that can autonomously recognize problems and find suitable solutions. I think we’ve come quite far with that.
If we now transfer this to AI, AI would be any machine that can autonomously recognize problems and find suitable solutions. I think we’ve come quite far with that.
The levels of AI and examples:
Level 0: No AI as we understand it would be machines that merely follow instructions. A computer that displays text on a screen that a human enters via a keyboard performs calculations on its own, but these calculations are not used to solve problems. These calculations are rather an end in themselves, so that the machine can perform its activities at all.
Level Ia) If we were to extend the previous example to include a spell checker, we could assume AI Level I. Because this already requires that the computer independently compares the text with a data set and determines whether there is a match. The AI can distinguish between correct and incorrect and return the result to us. This is a first autonomous problem solving, according to a given set of rules. Today, we do not consider this to be particularly intelligent, but 60 years ago we would probably have assessed it differently.
Level I b) While the spell checker only requires the computer to compare its data with an existing database, a search engine, such as the well-known Internet search by Google, has more advanced capabilities. However, this does not mean the execution of the search itself, but the filling of the database. While the database for the spell check of a word processing program is manually filled with data, Google searches with its own machines all computers that are connected to the Internet for data, evaluates them and stores them in its gigantic databases. This requires the search robots to recognize, sort, compare and store web servers, data patterns, content types,… etc.
This is where the difference lies, because now, for the first time, the machine has to recognize patterns itself. But we still stay in level I, because in this case pattern recognition happens according to logical rules (algorithms). Programmers have previously told the computer how it recognizes that a picture is a picture and a text document is a text document. Based on predefined criteria, the computer assigns the found document to a category under which it stores it. The whole thing seems quite intelligent, but is still based on purely logical processes. It is absolutely predictable (even if only with great effort) what the machine will do.
Further use cases are expert systems, which e.g. make possible diagnoses on the basis of determined symptoms, or navigation systems, which can calculate and optimize routes according to given algorithms, etc…
AI of this level enables computers to collect and process large amounts of data according to predefined rules.
Level II a) Now we have arrived at the point where AI really begins according to the current understanding. The basis for the knowledge of an AI of level II is no longer determined by logical rules (even if the results can be logically comprehensible), it is now based on learning and experience. The whole thing is made possible by artificial neural networks. They originate from the idea to reproduce the human brain in computers in order to break the superiority of humans over machines in experiential learning. Limited to individual areas, this already works wonderfully.
Image recognition is one such use case. There is no logical rule according to which we recognize, assign and evaluate images, but we have learned it from an early age, first from our parents and later through more and more images that we were able to compare and differentiate better and better. We can now also create our own images in our head and, if necessary, put them on paper. We cannot explain to a toddler what a beautiful picture is, we can only assign single pictures to the term beautiful and based on experience the child can later take over and further develop the knowledge about pictures.
Considerations of artificial neural networks have also been around for decades now, but only modern computer chips and memory are efficient enough to use neural networks for practical use cases. Add to that the progress of research and the opportunities that have been created by the proliferation of the Internet.
AI of this level has the potential to pass a Turing test, which has already been achieved in some areas. In any case, it is an application area where we will see a lot of new things in the coming years.
Level II b) Even if many already speak of Deep Learning in AI of level IIa, I still see a conceptual difference here, which, however, does not yet have any particular weight as of 2018 because the practical application is still in its infancy. Deep learning, that is, optimizing the learning of artificial neural networks, enables machines to continuously relearn learning. The differences are quite technical, so I won’t go into them in this article, but they will definitely become more relevant in the future.
Level III will occur when the singularity is reached and computers not only solve problems autonomously, but also autonomously decide which problems to deal with and which tasks to perform, e.g. as part of a control unit of a robot.
This sounds like a bit of science fiction, but it is probably not that far away.
To avoid any unpleasant surprises then, it is important that we have an open discussion around AI and steer the future in the right direction while we have influence over it.
Delimitations:
According to our considerations, artificial intelligence is limited to solving problems in the sense of making decisions, or providing selected information. However, it does not tell us what to do with it, nor does it perform any actions.
Robots: Robots are not AI. Robots are machines that perform activities. Their control can rely on artificial intelligence, they can also be controlled by humans, or they can execute very simple program instructions.
Chatbots: Behind this term is the fiction that you communicate with a robot (bot) via a messenger service (chat). Often, however, these are merely chats with web servers on which simple question-answer schemes are stored. However, there are also chatbots that access qualified expert systems or interact with level II AI.
Originally published in German at medium.com on Jan 7, 2018.
