In this article, we will introduce the types of AI. Have you ever wondered about the difference between the AI depicted in science fiction works and the AI that is familiar to us today? In fact, the two AIs are completely different.
In this article, in addition to introducing the types of AI, we will also touch on the types of AI that are currently in practical use and the types of AI that are expected to be realized in the future.
Also, in the second half of this article, I will introduce various problems that the type of AI suggests, so please read to the end.
Table of Contents
- AI (artificial intelligence) can be broadly divided into two types.
- Types of AI classified by task width
- Types of AI classified by level of intelligence
- Types of AI analysis technology (weak AI, specialized AI)
- machine learning
- deep learning
- Various problems indicated by the types of AI (strong AI, general-purpose AI)
- Is general-purpose AI possible?
- General-purpose AI as an extension of specialized AI
- General-purpose AI with a biological approach
- Is Strong AI Possible?
- A collection of analytical techniques cannot be said to be strong AI
- Strong AI Thought Experiment
- Achieving strong AI faces a mountain of technical challenges
- in conclusion
AI (artificial intelligence) can be broadly divided into two types.
There are several categories of AI depending on its nature and usage. In this chapter, we will introduce the types of AI classified by “task width” and “intelligence level”.
Types of AI classified by task width
First, let’s look at the types of AI classified by task width.
AI that can perform one task or a very limited range of tasks is called ” specialized AI “, while AI that can perform multiple tasks is called “general-purpose AI”.
I will introduce each of them.
“ Specialized AI ” is AI that specializes in performing a specific task. Most of the current AI belongs to the category of specialized AI.
For example, specialized AI includes AI for playing games such as shogi, chess, and Go, AI for image recognition and voice recognition, and AI for autonomous driving.
“General-purpose AI” is AI that can handle a wide variety of tasks that are not limited to specific tasks. In other words, it refers to AI that can behave as well as or better than humans.
Currently, there are few practical examples of general-purpose AI, and even the few examples are far from “human-like” AI.
Types of AI classified by level of intelligence
Next, AI can be classified into two types, ” weak AI ” and ” strong AI “, according to the level of intelligence . This categorization is often used in the debate over whether AI can acquire “human-like” reasoning and task abilities.
American philosopher John Searle called AI, which does not have a human-like mind and simply solves problems according to a program, ” weak AI .”
All currently prevalent AIs belong to this weak AI category. Weak AI cannot be said to be truly intelligent.
It is true that AI for shogi and go seems to far surpass human intelligence, but that AI is just processing according to pre-inputted algorithms.
On the other hand, ” strong AI ” is similar to the image of AI that is often talked about in science fiction works. Specifically, AI that has human-like intelligence and can solve problems autonomously can be said to be strong AI.
It’s no surprise that such a strong AI harbors a sense of self. Because such a strong AI has true human-like intelligence.
However, strong AI has not yet been realized. Searle, who devised this division, similarly thought strong AI was unfeasible. This issue will be explored in more detail in a later chapter.
Types of AI analysis technology (weak AI, specialized AI)
As mentioned above, all of the AI currently in practical use belong to the category of “ specialized AI & weak AI ”.
Each AI employs different kinds of analytical techniques to specialize in specific tasks. This chapter introduces the types of such AI analysis techniques.
Machine learning is an analysis technology that discovers data patterns by having AI learn data and uses the learning results to perform various tasks.
There are three main types of machine learning: supervised learning , unsupervised learning , and reinforcement learning .
“ Supervised learning ” is a method of learning data after giving correct answers to AI. This learning process is called “supervised” learning because it can be modeled on the teacher-student relationship.
Typical examples of supervised learning are regression and classification problems. Regression problems deal with continuous values, while classification problems deal with discrete values. For example, a program that predicts sales belongs to the regression problem, and a program that determines spam mail belongs to the classification problem.
Supervised learning is a form of machine learning that has been used in many situations for a long time because of its simple mechanism and algorithms.
” Unsupervised learning ” is a method of learning data without giving correct answers to AI. Find features in given data unsupervised.
Typical processes in unsupervised learning are clustering and dimensionality reduction. Clustering is used to divide (group) data into clusters with similar characteristics. Dimensionality reduction is used to reduce excess features in data to prevent overfitting.
Reinforcement learning is a method of learning so that an agent selects the action that can obtain the maximum reward in a certain environment.
Reinforcement learning has become famous for its AI successes in Go and Shogi. Reinforcement learning is also used in robot control and autonomous driving.
Deep learning is an AI method that enables much higher processing power than traditional machine learning methods. Deep learning analyzes huge data groups called big data.
Deep learning learns based on a model that mimics the neural circuits of the brain called a neural network. The neural network is made up of units called perceptrons.
Deep learning has revolutionized AI applications such as image recognition, speech recognition, and natural language processing. With conventional machine learning methods, it was impossible to analyze and process image data, voice data, and languages used by humans (natural language) such as English and Japanese.
However, the establishment of deep learning technology has made it possible to process this huge amount of data in a short time.
Various problems indicated by the types of AI (strong AI, general-purpose AI)
So far, we have looked at the details of weak AI and specialized AI. Next, let’s turn our attention to strong AI and general-purpose AI. These types of AI suggest different problems that AI presents.
Is general-purpose AI possible?
“General-purpose AI” is AI that is not limited to specific tasks and can handle various tasks just like humans. Is such AI even possible? Also, what kind of approach is the research progressing from?
General-purpose AI as an extension of specialized AI
General-purpose AI, which seems to be feasible by accumulating current technologies, is general-purpose AI as an extension of specialized AI.
General-purpose AI is AI that is not limited to a specific task, but the simplest idea is that if you combine countless AIs that specialize in individual areas, it will become general-purpose AI.
This approach is realistic, but the question remains whether general-purpose AI created with this idea meets the definition of general-purpose AI in the first place.
General-purpose AI with a biological approach
The approach described above is what Ray Kurzweil, an authority on artificial intelligence, has named the engineering approach. Kurzweil also advocated the realization of general-purpose AI through a biological approach.
In the biological approach, a method is taken in which a machine is connected to the human brain and attempts are made to directly analyze and reproduce the processing processes of the human brain.
The neural network used in deep learning imitates the neural circuits of the human brain, but the model is just an extremely simplified version of the complex processing processes of the human brain.
A biological approach is a very direct reproduction of the human brain. Of course, this technique is not feasible with modern technology, and I don’t know if it will be feasible in the future.
There are also many ethical challenges, as this approach directly intervenes in the human brain, possibly creating something like a human clone.
Is Strong AI Possible?
Next, I will introduce the question of whether strong AI is possible. Philosopher John Searle, who advocated the weak /strong AI distinction, made this distinction to explain the impossibility of strong AI.
A collection of analytical techniques cannot be said to be strong AI
Strong AI is AI with human-like intelligence and autonomous problem-solving abilities. A strong AI has human-like “intention” and “self-awareness”.
Therefore, it is possible to think that a strong AI has a “mind”.
AI, which is an extension of current AI, may be able to process “pain” information in the future, but that AI does not feel “pain” in an essential sense.
The analytical techniques of today’s specialized AI, or the future general AI as a collection of analytical techniques, are not strong AI. So, is it possible to create an AI that feels “pain” in an essential sense?
Strong AI Thought Experiment
Searle showed the impossibility of strong AI in his “Chinese Room” thought experiment.
A person who cannot understand Chinese at all is locked in a room.
He has a complete rulebook for answering Chinese questions, that is, a rule that says, “If a Chinese question comes in ○○, answer in Chinese XX.” I have been given a book that is
A person who doesn’t know about it throws a question written in Chinese into the room from outside the room. Then, according to the rule book, the person in the room writes the answer in Chinese on a piece of paper and throws it back out of the room.
To people outside the room, it looks like people inside the room understand Chinese perfectly, but they really don’t.
Similarly, artificial intelligence that simply processes what is written in a program, even if it can perfectly process Chinese questions, cannot be said to understand Chinese in an essential sense. is.
Achieving strong AI faces a mountain of technical challenges
For these reasons, the realization of strong AI requires an AI with a mechanism that is completely different from conventional AI. It can be said that there is no strong AI on the extension of the current AI technology.
Searle also used the Chinese Room thought experiment to pose the question of human understanding and self-consciousness. Considering whether strong AI is feasible invites us to confront the open question of what the human mind is in the first place.
In this article, we talked about types of AI. He introduced the categories of specialized AI /general-purpose AI and weak AI /strong AI, and touched on the current technological status and challenges for each.
There is a big gap between the AI depicted in science fiction works (general-purpose AI/ strong AI ) and the AI currently in practical use ( specialized AI / weak AI ).
The types of AI suggest different problems that the gap suggests. We hope that this article will help you to deepen your understanding of the question, “What is AI?”