This article is reproduced from the public number of the door venture,Original address

As we explore the concepts and algorithms behind artificial intelligence and machine learning, we come into contact with a range of technical terms and core concepts related to this field. Understanding these terms and concepts will help us better understand the development of this field and understand how data scientists and AI researchers are leading the way.

In this article, we will help you understand betterSupervised learning, unsupervised learning and reinforcement learningThe definition of the definition and the connection between them and machine learning from a broader perspective. A deep understanding of their connotations will not only help you to be embarrassed in the literature in this field, but also guide you to sharply capture the development of AI and the advancement of technological progress.

Supervised learning, unsupervised learning, and intensive learning describe three different approaches to machine processing and data learning. People use different learning methods to condense knowledge from data according to different data and tasks, thus helping humans in production and life. .

Supervised learning

In order to let more friends understand the core concepts of machine learning, we first give a concise definition of concise supervision and learning:Train a model with input and output and map the input to the output.We already know the input and input before we start training. Our task builds a model that maps the input to the output accurately. When the new value is entered into the model, the corresponding output can be predicted.

In the process, the machine continuously passes the training input to guide the continuous improvement of the algorithm. If the output is not correct, the error between this erroneous result and the expected correct result will be passed back to the model as a correction signal, correcting the model improvement. The well-known backpropagation algorithm in deep learning fundamentally also propagates the error backwards to guide the model improvement.

At present, supervised learning occupies the vast majority of machine learning algorithms. The algorithm combines the input variable x with the output variable y to create many applications with far-reaching effects. To understand supervised learning, we need to grasp the following points. All the algorithms, inputs, outputs, and scenes in supervised learning are provided by humans. Dividing the supervised learning problem into two categories will better help us understand the meaning of supervised learning.

Classification: The goal of a classification problem is to predict the category to which the sample belongs by input variables. For example, prediction problems for plant varieties, customer age, and preferences can be attributed to classification problems. The most used model in this field is the support vector machine, which is used to generate decision boundaries for linear classification. With the development of deep learning, many image signal-based classification problems are increasingly completed using convolutional neural networks.

Regression: It is mainly used to predict the real value of a variable, and its output is not a classification result but an actual value. The most common examples are market price forecasts, precipitation forecasts, and more. Regression models are constructed primarily through linear regression, polynomial regression, and nuclear methods.

Unsupervised learning

After learning about supervised learning, let's take a look at the unsupervised learning method. This method is not as widely used as supervised learning. The current practical application is not as common as supervised learning. But this unique methodology gives a lot of inspiration and possibilities for the future direction of machine learning. Perhaps unsupervised learning allows us to evolve from “what the church machine does” to let the machine “learn to do what it does”.

Different from supervised learning,Unsupervised learning does not require a complete input and output data set, and the output of the system is often uncertain.It is mainly used to explore the patterns and distributions implied in the data. Unsupervised learning has the ability to interpret data and seek solutions from it. By entering data and algorithms into the machine, you can find patterns and information that are otherwise inaccessible.

Let us use a simple example to understand unsupervised learning. Imagine that we have a collection of photos that contain geometric shapes of different colors. Here the computer is facing a picture without any mark, it does not know the color and shape of the geometry, it only sees a single photo. But by inputting the data into the unsupervised learning model, the algorithm can try to understand the content in the graph, and similar objects in the graph can be grouped together by correlation and feature. Ideally, it can aggregate geometric shapes of different shapes and colors into different categories. Feature extraction and labeling are all done by the machine itself.

But like human learning, machines make mistakes. The ability of the machine can be improved by identifying errors and learning from mistakes.

Reinforcement learning

Reinforcement learning is an important part of machine learning and has made tremendous contributions to the new direction of machine learning. Reinforcement learning breaks through unsupervised learning and gives a whole new idea of ​​how machines and software can achieve optimal results.It establishes a strong link between how to optimize the performance of the subject and how to optimize it.Help the machine improve its behavior and algorithms through feedback from reward functions.

However, reinforcement learning is not simple in practice. People use a variety of algorithms to achieve reinforcement learning. In simple terms, reinforcement learning requires the machine to behave in the current state to get the best results.

In reinforcement learning, the subject interacts with the environment through behavior, and the environment uses the reward function to help the algorithm adjust the strategy function that makes the behavior decision. Thus, a behavioral strategy with excellent performance is obtained in a continuous cycle. It is ideal for scenarios such as training control algorithms and game AI.

Finally, let us discuss after understanding the concepts of these three.The similarities and differences between supervised learning, unsupervised learning and intensive learning:

1. Supervised learning vs reinforcement learning

In the supervised learning, the corresponding input and output data plays the role of supervision, and the knowledge contained in it is given to the model through training. The model continuously learns through the correction signal of the data and finally forms an algorithm that can better understand the data and accurately predict. It is not necessary to supervise data in reinforcement learning. The subject can try many methods and performances and adjust through interaction with the environment. We can imagine the example of Go. We need to perform many operations before the final outcome is announced. There are many possibilities for each operation. It is very complicated to establish the knowledge of supervised learning for this task. The computer can build its own understanding of the game based on feedback received after interacting with the environment.

When the machine starts learning, she can constantly enrich her experience and improve performance. This may be the biggest difference from supervised learning. Although both models establish some kind of input-to-output mapping, reinforcement learning uses the reward function to help the system continuously improve the model.

2. Intensive Learning vs Unsupervised Learning

Reinforcement learning essentially links input and input through a mapping structure, but unsupervised learning does not have any links between input and output. In unsupervised learning, the main task of the machine is to identify patterns of data types rather than to establish mapping relationships. If we want to build a user news recommendation system, reinforcement learning can be continuously improved through user feedback, and a user-like news type map can be established to achieve more accurate recommendations. For unsupervised learning, the user's preferences can be summarized from the articles read by the user, and appropriate topics can be recommended for the user.