It depends on how artificial intelligence (AI) is defined, but according to "What AI can and cannot do" [^ 1]
"Technology that allows computers to perform intelligent work"
It seems that. Computers can only act as calculators by themselves, but intellectual work cannot be done by calculations alone. Today, only weak AI (artificial intelligence that is good at a specific task) is made, and the center of research is also weak AI. However, as seen in AlphaGo and the University of Tokyo Robo, AI that exceeds human ability is beginning to appear, and although it is imperfect, it attracts people. In addition, AI is emerging in advertising and political activities, and it is becoming impossible to avoid it in daily life, such as displaying fake videos and personalized advertisements. If a strong AI is completed, when the computer can speak, express, and do various things by itself, it will overcome the singularity and the game "Detroit: Become Human" It may become a reality that AI attacks humans by ignoring human intentions or running away like the movie "I, Robot" (although it is extreme). It is also said to be a 2045 problem. At present, AI is just an algorithm, so in that sense, I think it is unlikely that it will exceed the singularity.
The story is derailed, but there are four main types of artificial intelligence.
The above three are collectively called machine learning. To put it simply, machine learning is an algorithm that automatically learns conditional branching. Predictive models are not created by humans, but are automatically generated. Still, the values of certain parameters must be determined by humans. Such parameters are called __hyperparameters __. Also, as a subset of machine learning, there is __deep learning __.
Although it is generally regarded as "AI = deep learning", AI is not limited to deep learning. The reason for its attention is that deep learning mimics nerve cell neurons and is closer to the structure of the human brain. However, when recognizing objects and characters, there is a big difference between human recognition and computer recognition. In that sense, it would be impossible to compose the human brain.
By the way, AlphaGo uses Deep Reinforcement Learning, which is a combination of deep learning and reinforcement learning.
Now, let's take a quick look at some techniques for each of machine learning.
Supervised learning is a technology that learns a model based on teacher data (input data + correct answer label) and predicts and classifies unknown data. For example, is it an image that a teacher teaches a student how to solve an example, and then the student who learns the solution solves an exercise? The goal is to increase the correct answer rate for the student. However, if you teach too much, you will develop strange habits and will not think for yourself, and as a result you will not be able to solve simple problems that you have solved before. This condition is called overfitting. Overfitting is, so to speak, overfitting the training data. In other words, the image is that students can only solve the problem as taught, and they cannot solve it immediately by changing the wording and numerical values of the problem sentences. Therefore, it is necessary to create a model that minimizes the error while adjusting the parameters so as not to cause overfitting. There are two types of supervised learning: regression and classification problems. Let's first look at regression.
(Following the outline of supervised learning ...)
[^ 1]: Koji Fujimoto and Kazutomo Shibahara "What AI can and cannot do" Nihon Hyoronsha, 2019
Recommended Posts