Why is artificial intelligence being talked about?
May 28, 2017：Computer Science
Recently, artificial intelligence is often taken up in the news. For example, automatic translation by Google, topics of AlphaGo and automatic driving. So, why is artificial intelligence recently talked about? I will talk about deep learning (deep learning) where artificial intelligence became popular.
Image processing so far
I think that the most obvious example is image processing to understand the development of research by artificial intelligence, so I will show you how image processing was changed by deep learning. In conventional image processing, programmers saw images and built their own algorithms. For example, we use algorithms to find out whether there is a person’s eyes, nose or mouth in the identification of whether there is a person in the image, and in the presence of these, we judge whether it is a person’s image or not. As you can see from this example, solving the question “Is this image a person?” Program is very painstaking. Even with the algorithm just before, even if there are eyes, nose, mouth, you can not say people. Maybe it is a cat or a dog.
As mentioned above, it is very difficult task for us to make an algorithm that we can immediately distinguish from “being a person” as an algorithm. Researchers and engineers have studied new algorithms to raise the accuracy of this difficult task as much as possible. I never dreamed that this hardship would be automatically overcome by the calculator.
Machine learning by neural network
Dr. Kunihiko Fukushima, a Japanese researcher in 1979, was doing research to model nerve cells in human eyes. Nerve cells will fire when the input signal exceeds the threshold and will output. We model this as follows.
The [latex]f[/latex] that comes out here is a function, and think that it outputs if it becomes larger than 0. And [latex] w [/latex] and [latex]b[/latex] are constants called weights and biases. Giving appropriate inputs to this nerve cell will produce a constant output. Here, the problem is “How to decide weights and biases”.
To decide weights and biases, prepare teacher data. Teacher data is a data set that defines that for an input [latex]x[/latex], the output is [latex]y[/latex]. For example, teacher data is one that inputs a number image and outputs the number written in the image. Based on this teacher data, we decide the weight and bias. This process is called “learning”.
Once you’ve learned, just give the input and output the result. For example, if you give a number image with 3, the number 3 will appear. The good point of the neural network is that it only has to give teacher data. In conventional image processing, engineers built their own algorithms. However, neural networks automatically form algorithms from teacher data.
At first glance, the neural network seems to be OK, but there was a problem that the accuracy was worse than the algorithm created by a technician. Therefore, not much attention was gathered and years passed. Thanks to that, the specifications of the computer increased steadily as a result of micromachining.
Complex neural network exceeds human’s algorithm
With the improvement of computer specifications, it became possible to learn quickly even if the neural network became complicated. As a result, we developed into “deep learning (deep learning)” that layered neural networks. We will lay nerve cell networks in layers as follows. Then, by giving teacher data, learning determines the weight and bias of each nerve cell. However, just as the number of neurons increases, it does not change anything from the past. However, we found that increasing the number of neurons improves the accuracy, and when I noticed, I surpassed the algorithm that people made.
The improvement of the number of neurons by improving the computing power has dramatically improved the performance of artificial intelligence. The algorithms that people have manually created have passed beyond all fields, and deep learning has gained attention. Neural network technology is simple, so it can be used in various fields such as audio processing as well as image processing. Therefore, deep learning will be applied everywhere in the future, I think that it will become a convenient world.
What is applied to
Finally, I will describe what deep learning is applied and what can be done. I hope you find it helpful.
- Automatic recognition of goods inside the store (eg recognition of bread)
- Face authentication login of personal computer or smartphone
- The key of home or company by face authentication
- Automatic translation (Google, Facebook)
- Tracking people from videos (Amazon GO !?)
- voice recognition
- Character recognition (OCR)
- Recognition of what is shown in the picture