Let’s start with some basics. For those of you who are now rolling your eyes, you can start right away with my second post. 😉 For those who think of artificial intelligence as a sci-fi topic, which is far far away from our daily life: you are exactly right here.
There is a huge amount of content on artificial intelligence and I really don’t want to just repeat what many others have written before. That’s why I will link to other smart guys whenever necessary.
Artificial Intelligence: The Three Stations of Evolution
So let’s clear things up. First, stop thinking of robots. A robot is a container for AI — sometimes mimicking the human form, sometimes not — but the AI itself is the computer inside the robot. AI is the brain, and the robot is its body — if it even has a body. For example, the software and data behind Siri is AI, the woman’s voice we hear is a personification of that AI, and there’s no robot involved at all. 1
“Artificial Intelligence (AI) is a science and a set of computational technologies that are inspired by—but typically operate quite differently from—the ways people use their nervous systems and bodies to sense, learn, reason, and take action.” 2
To put it simple Artificial Intelligence is the automation of intelligent behavior. It is the ability of machines to solve tasks that are normally associated with the higher intellectual abilities of a person. But Artificial intelligence is not always the same. There are three steps in the development of artificial intelligence:
- Artificial Narrow Intelligence (ANI) is also known as Weak AI. It is specialized at one particular area or one particular task. Thus, an ANI that can recognize language isn’t able to draw a picture. An ANI that is able to play chess can’t write a letter.
- Artificial General Intelligence (AGI) is known as Strong AI. AGI refers to a computer that has the intelligence of a human across the board. A machine that can perform any intellectual task that a human being can. That’s why it is also called human level AI.
- Artificial Super Intelligence (ASI): A Super Intelligence is an intellect that far exceeds human cognitive performance in almost all areas.
Machine Learning, Deep Learning, Neural Networks
One approach to achieving artificial intelligence is through Machine Learning, systems that learn from experiences to find patterns in a set of data.. It describes a mathematical technique that enables a computer system to generate knowledge – in other words, to learn.
The intelligent system takes data, analyzes it and tries to recognize patterns and regularities. It learns these patterns from the dataset and it extracts rules from them. The more data is passed through the system, the greater the knowledge gets and the more precise the patterns become that the system identifies. This way, it learns to improve its behavior. And the better the behavior, the better the predictions will be. In other words, Machine Learning is about creating algorithms that learn patterns from data and make predictions based on them.
Besides the ability to learn, another integral part of artificial intelligence is the effective handling of uncertainty. It has the ability to extract useful terms from sense data and internal states and to use them in flexible combinations for logical and intuitive thinking.
A promising type of machine learning is Deep Learning, especially a concept known as deep neural networks. The idea of a neural network is a system structured like the human brain. A comprehensive explanation of Machine Learning, Deep Learning, Neural Networks and how it all works can be found here:
Artificial neural networks are distinguished according to the degree of human influence. While supervised learning is the most widespread form of learning today, unsupervised learning will become increasingly important in the coming years.
- Supervised Learning: In supervised learning, Data Scientists act as guides for the algorithm. They check the output of the input and the weighting. If necessary, a statistical procedure is used to correct errors. Simplified: they teach the machine what’s the right result. Probably the best-known example of this is AlphaGo. Data Scientists taught the system the rules of the game and trained it with millions of positions from games between strong human players. Ultimately, AlphaGo playedGo better than the best human player in the world. AlphaGo Zero has now learned the game from scratch, just based on it’s rules.
- Unsupervised Learning: In the case of unsupervised learning, the systems independently recognize how to deal with the raw data. Through unsupervised learning, AlphaGo Zero taught itself to play the game within three days without human influence,playing against itself. The result: Artificial intelligence has not only taught itself new knowledge, it has also designed unconventional methods and taken creative new steps. This will ultimately enable people to ask questions, solve problems, generate new ideas about patterns that they have not come up with yet.
- Reinforcement Learning: Neural networks are not only able to recognize patterns, they can actively act as ‘artificial agents’ in their environment and explore them. The agents learn from the statistical calculation of the patterns what it is and what they have to do in this environment – an almost evolutionary learning process. The agent’s impulse to act is based on a reward model. He acts because certain kind of behavior is rewarded. A reward can for example be a real number. The system will therefore optimize everything to achieve this objective function. 3
The key aspect of Deep Learning is that the neural layers no longer have to be weighted by a programmer, but that this is done autonomously – only on the basis of the dataset that is processed through a learned procedure. Thus, there will be the basic functions of neurons (input, activation, output) that process incoming data independently and translate it into a mathematical function. This will solve problems that are inherently not programmable due to their inate complexity. 4
Artificial General Intelligence is on the horizon
In 1996, the intelligent system ‘Deep Blue’ developed by IBM beats the chess world champion Garry Kasparov,in 2011 IBM Watson wins the USA’s best Jeopardy players and in 20017Googles AlphaGo beats the world’s best go player and Libratus wins against four top poker pros in heads-up No-Limit Texas Hold’em.
Of course, these are examples particularly geared towards the media and the public, but nevertheless two things become clear: Firstly, intelligent algorithms are superior to people in more and more tasks and secondly, the time intervals between these events are becoming significantly shorter. The performance of artificial intelligence is increasing at an enormous rate.
However, currently we are still on the first step. Today’s AI applications are experts on specific areas. They are very good in playing chess, identifying cats in pictures, recognizing voices or driving cars. And in some of these areas they can already beat humans hands-down. But they have no true general intelligence. This means the AI is very intelligent for example in the domain of chess and it is absolutely useless in any other domain. So if you put a chess AI in a self driving car, it is not only incapable of driving that car, moreover it doesn’t even know what a car is and it doesn’t have any of the cognitive architecture to drive this car. At the moment there is only one being that has a general intelligence: it’s us.
But what about the future? Will computers become as intelligent as humans? Will they have a general intelligence? Yes, I’m pretty sure this will happen in our lifetime. It is very likely, that we will see the day when computers achieve the same general intelligence as humans. When will this happen? There is some debate about how soon AI will reach human-level intelligence. Looking at a survey of hundreds of scientists on when they believed we’d be more likely than not to have reached AGI and assuming that they know what they are doing, the following picture appears:
- The probability that AGI will be achieved by the year 2022 lies at 10%.
- The probability that AGI will be achieved by the year 2040 lies at 50%.
- The probability that AGI will be achieved by the year 2075 lies at 90%. 5
In other words: From 2045 the human race might no longer be the top species on earth! Machines will be able to perform any intellectual task that a human being can.
And as soon as we get to AGI, we’ll surpass it very quickly. The increasingly accelerated progress of artificial intelligence will not stop at the stage of human cognitive ability, but will simply overtake it on its way into the stratosphere. From then on we will have computers more intelligent than their human masters. And then things will truly get exciting.
The mathematician Irving John Good (1916-2009),chief statistician of the group around Alan Turing, cracked the German cipher machine Enigma during World War II. He wrote in 1965:
“A super-intelligent machine is defined as a machine that can far surpass any mental effort of any smart person. Since the construction of machines is such a mental effort, an super-intelligent machine could construct even better machines; undoubtedly an explosion of intelligence would then occur […]. The first super-intelligent machine is the last invention mankind ever has to make […].” 6