From this time, we will send you a lectured content.

Introduction to machine learning

What is artificial intelligence and machine learning?

Artificial Intelligence: According to Yutaka Matsuo of the Japanese Society for Artificial Intelligence "Artificially created, human-like intelligence, or the technology that makes it"

The details vary from expert to expert, but they define almost the same content.

Machine learning: analysis of specific data → learning → judgment → prediction It is a method that uses an algorithm that does

As shown in the figure below, deep learning is part of the classification of artificial intelligence and is more detailed.

Practical examples of artificial intelligence

Self-driving car
Lights that turn on automatically when you enter the room

Basically, it works semi-automatically in response to human commands.

Artificial intelligence and deep learning

Deep learning: A type of machine learning technology that mimics the brain of a living organism. It is a technology that uses a neural network.

By instantly extracting elements such as red, large, and bright from the image It is possible to realize processing far beyond humans.

Why is it attracting attention: Deep learning technology was developed in 2012 and has been higher than humans until now. The error rate of image analysis has been dramatically improved. Beyond humans in 2015 Because it came to be.

Reasons for machine learning attention

With a new technology called machine learning Analyzing large amounts of complex data at speeds far exceeding humans This is because it has become possible to cut the conventional cost and time.

In recent years, the processing speed of computers has improved, but

By reading patterns from a large amount of data, it is possible to acquire the required data.

Machine learning is one of the research subjects of "artificial intelligence" that has the same intelligence as human beings. In a wide range of fields such as image, audio, marketing, natural language, medical, etc. It is attracting attention because it is a technology that can be utilized.

What is machine learning?

A simple explanation of "machine learning" is "learning iteratively from data and finding the patterns hidden in it." Learn (create a database) elements and patterns such as squares, spheres, and shades, and from those learning contents We will detect the answer.

For example, in order for a computer to recognize apples, it needs to get a common pattern from a large number of apple photos.

And there are three main methods of machine learning to realize machine learning.

Supervised Learning
Unsupervised Learning
Reinforcement Learning

Symbol grounding problem

Even if you know the pattern that "apples are red and have a radius of about 5 cm" It misidentifies a 5 cm red ball as an apple. In this way, the problem is that characters (symbols) cannot correspond to the real thing.

Machine learning methods

Supervised learning

"Supervised learning" is one of the typical methods of machine learning. In this case, "teacher" refers to the "correct label attached to the data".

The mechanism is after learning with "learning data" and giving an answer using a method called "machine learning algorithm". Look at the correct answer "correct label" and match the answers.

The data in which "learning data" and "correct answer label" are combined is called "labeled learning data (teacher data)".

The "correct label" is the label that answers the data. "5" is attached as the correct answer label to the image data of the handwritten characters of data ①. And data ② is the image data of a vague horse, and "horse" is the correct label.

The technology for recognizing images in this way is called "image recognition," and is a specialty of deep learning.

On the other hand, data ③ is not an image but text, and the correct label is also the text "Natsume Soseki". This technique is called "natural language processing". In learning natural language processing, it is necessary to prepare a dataset for each language.

The flow of supervised learning is as follows.

1, Input multiple teacher data to the computer and learn the pattern model of the data 2, Input another teacher data to the trained model and verify the output. 3, Predict how new data will be classified from the learned model The basis of supervised learning is to be able to distinguish.

Classification problem

Data (1) to (3) that finally predict the category

Regression problem

A problem that predicts numerical values such as rent based on quantitative data such as data ④ is called a "regression problem".

Unsupervised learning

Whereas there is a teacher called the correct label in supervised learning There is no teacher in unsupervised learning. In unsupervised learning, the computer itself finds similarities and regularities from the given data. Learn the model.

Therefore, in the case of unsupervised learning, the feature is that there are no correct or incorrect answers.

You can see that there are three groups at the points in the figure. To make the computer recognize these three groups One of unsupervised learning We use a technique called clustering. Clustering allows data grouping.

Unsupervised learning is used to derive the laws of data and to group them. In actual analysis, multiple methods are often used in combination as shown below.

Clustering

A major feature is that the data can be grouped based on the similarity between the input data.

Principal component analysis

Principal component analysis is a technique used to summarize (dimension reduce) a wide variety of data. It has the advantage of making it easier to understand data trends and characteristics.

Association analysis

It is a method to find a rule of data such as "data that applies to one pattern also applies to another pattern". Association analysis is used for recommendations such as "People who see this product are buying this product" that you can see at online shops.

In this way, unsupervised learning is used to derive and group the laws of data. In actual analysis, multiple methods are often used in combination.

Reinforcement learning

Reinforcement learning is a way to maximize profits and is autonomous machine learning that does not require correct labels or large amounts of data. Recently, it is often combined with deep learning, such as games and other competition programs. It is used for control programs of walking robots.

The following keywords appear in reinforcement learning.

Agent
Environment
Action
Reward

To explain, in reinforcement learning, an agent (acting subject) acquires and observes a given environment and takes action. Then, we will maximize the value of the reward (result) obtained by the change of the environment due to the behavior. Agents learn autonomously by making repeated decisions through trial and error in order to obtain higher rewards.

As a legend, you can learn and develop better results in board games such as Go.