12 April 2023

Data engineering

Types of Artificial Intelligence - a general overview of a formidable technology

15 minutes reading

Types of Artificial Intelligence - a general overview of a formidable technology

We live in a world that even a couple of decades ago only science fiction writers could imagine. The most amazing innovations quickly become commonplace and normal. However, if there is still one that doesn’t fail to impress, that would be artificial intelligence, as proven by the recent burst of interest in ChatGPT, a new language model application with advanced AI features that can be used to build chatbots. 

We encounter AI everyday, not only in chatbots, but also in voice-activated personal assistants, self-driving cars, robot vacuum cleaners, image generation software, and many, many other instances. It might appear that we are very close to the end of the road that leads to designing an AI that will be smarter than human beings, and that makes some of us wary of what that achievement might mean for humanity.

If you look at the types of artificial intelligence more closely though, it becomes clear that we have barely scratched the surface. In this article we will review the types of AI and its stages of evolution so that we can understand its potential and future development.

An image of a data engineer analyzing data. Text says: "Make your data truly valuable"

Some AI history

Artificial intelligence might be a modern term, but people have actually imagined intelligent machines that are as smart or even smarter than the human brain centuries ago. Even one of the myths of Ancient Greece described Talos, a giant bronze automaton that protected the isle of Crete from pirates. 

One of the greatest theoreticians of computer science, Alan Turing, designed a special test, that he called an imitation game, in 1950 to test the ability of a machine to pass as a human in a conversation. The Turing test has become an important concept in AI philosophy, even though the term itself wasn’t coined until 1956 when John McCarthy, together with several other computer scientists, used the term ‘artificial intelligence’ in a conference proposal, which essentially started the field of AI. 

Today, when we speak of artificial intelligence, we usually imagine a smart machine that can appear to think and behave like humans do, or, in a wider sense, the branch of computer science that deals with creating such AI machines.

In one of our articles, “AI and Machine Learning for networks” we described how AI can help solve some network issues.

Types of AI

To better understand the potential of artificial intelligence and what it can bring us in future, it is important to get to know each type of AI in more detail. There are four types of AI systems, based on their functionality. 

Reactive Machines

A reactive machine is the most basic type of artificial intelligence. It has been decades since the first reactive AI systems were created. Reactive machines don’t have any functionality based on memory. They don’t store memories, so they can’t use them as a basis for decision making and hence are unable to learn. A reactive AI only reacts to the stimuli available in the moment. It reviews the scenario that is right in front of it and chooses the best way to act based on that information only.

A typical example, most often used to showcase a reactive machine, is the supercomputer named DeepBlue, which was built by IBM in the 1990s. It is famous for winning the 1997 chess match against grandmaster Garry Kasparov. DeepBlue was built to perceive the world directly and decide how to act accordingly, so it doesn’t try to consider a wide range of possible moves. It views the board, knows how the chess pieces move, and simply makes the next best move that is available right now.

Another reactive machine, AlphaGo, developed by Google, uses a neural network, so its analysis method is a bit more complicated compared to DeepBlue, but it also doesn’t try to predict all the potential moves that can happen in future. It was able to win multiple matches with some of the greatest Go players in the world.

There are purely reactive machines in your everyday life that perform certain basic functions in real time, like reviewing your previous streaming service searches and recommending new movies based on the preferences detected. They are great in situations where repetitive tasks, like SPAM filtering, need to be performed, because AIs can’t get bored or sleepy. They remain reliable and always fulfill the set task exceptionally well, but if you try to make reactive machines act beyond the specific situations for which they are trained, it is likely they will fail.

Limited Memory

AI with limited memory is another type of artificial intelligence we have already managed to build. These AI systems have all the functionality of reactive machines, but in addition to that there is also an ability to store memories for a limited period of time, which is typically pretty short. These memories are used by the AI for learning and improvement of its capabilities.

Limited memory machines need to be trained so that they can build a reference model which will be used in the future to solve different problems. The training can involve reviewing lots of specific data, for example, images for an AI that is being trained for image processing and recognition. The more labeled images that are used in the training, the better the AI becomes in recognizing new images and labeling them itself.

Another way limited memory AIs can obtain knowledge is by observing real-life events. That is how self-driving cars operate. They collect information about the cars around them and the way they move, and then combine it with the pre-programmed knowledge about traffic lights, road signs etc. when they need to make a driving decision. However, the car doesn’t save this information permanently, like a human driver could do.

Most present day AI systems are limited memory systems. Two other types of artificial intelligence are only conceptual at this point, but are they as far away as we believed earlier?

Theory of Mind

Although limited memory AI can imitate human-like behavior, it is unable to understand human feelings and emotions and react to them. It doesn’t matter what tone of voice, angry or sad, you use to address your voice-activated personal assistant, the reply will be based only on the meaning of the words you say. AI can’t see you are having a rough day and suggest where you can get a nice spa treatment to relax. However, the next level of AI will have emotional intelligence that will allow the machines to have social interactions, just like humans.

Theory of mind AI received its name from a psychological term that means the ability to understand how mental state can impact the behavior of yourself and others. This advanced AI won’t make racist or sexist remarks like ChatGPT inadvertently can, because it will understand that it might hurt the feelings of a human being.

Although this particular type of AI remains a work in progress, some attempts to increase the ability of AI to understand human feelings and respond to them have already been made. One example, a robot called Kismet, built by Professor Cynthia Breazeal from MIT, dates back to the late 1990s. It could read human facial expressions and recognize the expressed emotions, then reproduce them with its own face that had the same structure as a human one. 

Another attempt is a more advanced robot that looks like a human, developed recently by Hanson Robotics. Sophia is distinctly humanoid and can change the expression on her face in response to visual stimuli processed by the cameras in her eyes.

So, it looks like there are some advancements in developing the theory of mind level of AI, but a machine that can fully perceive emotional signals of a human being hasn’t been created yet.

Self-Aware AI

The last type is self-aware AI, such an AI wouldn’t only understand the feelings of others but be aware of itself as an individual. Self-aware AI will be conscious, capable of having its own ideas and wishes, but be much smarter than a human brain, and it can even potentially be superior to humans in certain areas. 

The advancement of technology to such an extent might turn out as a huge achievement for the whole human civilization. Even though some people are biased against self-aware AIs, based on their depictions in movies and books, their worries might be a bit premature. Anyway, taking into account that even developing theory of mind AI is still a long way away, there is a possibility that decades or even centuries separate us from the AI point of singularity, which is how the self-awareness phase of AI development is also known.

AI evolution stages

There is another widespread classification of AIs, based on AI capabilities at different stages of AI evolution. These stages are sometimes combined with the AI types, but, in fact, they categorize AI from a different perspective.

Narrow AI

Artificial narrow intelligence, also known as narrow AI or weak AI, is the kind of AI that fulfills a specific function but is unable to perform outside of predefined conditions. All the AIs that humanity has managed to build by now, both reactive and limited memory, are classified as narrow AI. 

General AI

The next stage in AI evolution will be artificial general intelligence, or strong AI. This AI is characterized by being able to learn and function just like a human being. They will be able to train independently and teach themselves, improving their own abilities. AI researchers are actively working on achieving this, but it is clear that the current advancement is far from the target. For example, one of the strongest supercomputers available at the moment, K by Fujitsu, was able to simulate one second of human neural activity but spent 40 minutes on that task.

Super AI

The ultimate evolved AI that will surpass human intelligence in its capabilities will be artificial superintelligence. This AI will be able to think faster, make better decisions, but also will be self-aware and potentially will have some desires of its own. The good news is that we have a lot of time to think about how to bring super AI into existence and learn how it can coexist with humans. It is obvious that technological advancement is far from the level that is required to build super AI, since we haven’t even gotten to general AI yet, even though the AI field in general is definitely one of the most popular among researchers.

Artificial intelligence is an amazing technological achievement that proves we are indeed living in the 21st century. Smart devices with AI capabilities surround us in our daily lives and make them much more pleasant. Thanks to AI, science research moves faster than ever before. We are clearly on the verge of many amazing breakthroughs.

Conclusion

Now that you know more about the currently available types of AI, it should become clear we are still very far away from artificial superintelligence. AI research, focused on learning and making decisions based on memories and past experience for development purposes, is bound to bring many new exciting theoretical discoveries. That is a great benefit on top of all the other practical enhancements that our lives will have thanks to AI becoming more and more widely available.

Volha

Volha Duplenka

Technical Writer