The essence of machine learning is that computers learn knowledge from data. This process is very similar to the process of human learning. Therefore, machine learning can effectively help people solve problems.
How does human knowledge come from? Aeschylus said: "Memory is the mother of all wisdom" ("Prometheus Bound"). Indeed, without memory, there is no wisdom. Most of the knowledge we have is obtained from memory. So the learning process is destined to be hard. No matter how many memory methods and shortcuts you have, you ca n’t break through the limitations of the human brain itself, and you have to remember them bit by bit. My alma mater, the Juvenile Class of the University of Science and Technology of China, is the most well-known place for cultivating juvenile genius. As far as I know, the youngest student admitted to the Juvenile Class is 11 years old, which is almost the limit of the minimum age. No matter how talented, seven or eight years of knowledge accumulation cannot be skipped, there is no absolute shortcut.
The English version of Aeschylus's famous quote is: Memory is the mother of all wisdom . The memory in a computer is called Memory, and memory is used to store data. In the world of computers, this famous saying should say: "Data is the mother of all wisdom." A computer has a huge advantage over humans. Its "memory" (that is, the process of collecting data) requires a very short time, far shorter than the time required by humans. From this perspective, if the computer replicates the learning process of human beings, then machine learning is more powerful than humans is a matter of course. In fact, this is already the case on many issues.
In the field of machine learning, data is often more important than algorithms, and data determines the upper limit of machine learning capabilities. Data / memory is the foundation of knowledge, but obviously they are not equal to knowledge. First, we need to "decode" the data, otherwise they are just a bunch of "01" strings; second, we need to identify the information that is useful to us.
How do people obtain knowledge from memorized data? Use the very simple example of "literacy" to illustrate. There are about 100,000 Chinese characters and about 3,500 commonly used Chinese characters. Before we have learned how to recognize characters, even if you can remember what they look like, they are just "characters" in your eyes. According to the definition of information entropy ( https://baike.baidu.com/item/%E4%BF%A1%E6%81%AF%E7%86%B5 ), if you are given a word, there is no difference in your eyes , Then it has only one possible value: "word", and the probability is 1.0, we can calculate this amount of information.
Therefore, it is useless to remember, if you can not recognize the meaning of each word, the amount of information you have is still 0. Consider only commonly used Chinese characters, assuming that you have mastered all 3500 Chinese characters, and then assume that the probabilities of these characters are the same, so the amount of information you have is as follows:
At the beginning of mankind, there is no good or evil. The world is chaotic, and we know nothing about it. "Heaven and earth are not separated yet, the universe is in chaos. There is a giant named Pangu who has been sleeping for 18,000 years in this chaos. One day, Pangu woke up suddenly. When he saw the darkness around him, he swept away Take a big axe and slash towards the darkness in front of you. Just listening to a loud noise, the chaotic things gradually separated. The light and clear things slowly rose into the sky; the heavy and turbid things slowly fell , Turned into the earth. " Pangu opened the earth , is the beginning of human beings; we know words and break words, discern right from wrong, is the beginning of wisdom.
One of the most important types of problems in machine learning is classification, which makes computers intelligent. The concept of classification is easy to understand. Right and wrong, good and evil, beauty and ugliness are all classifications, and the mathematical model of classification is also easy to understand, as shown below.
What the machine needs to learn is a "segment line", which can distinguish the points belonging to different categories on the plane.
I have always felt that even if some theories are more difficult, the concept of machine learning and the problems to be solved are well understood. If more people can understand some relevant knowledge, AI will no longer be a black box to the public, which will reduce a lot. Unnecessary social issues.
Examples of application of classification problems
- Spam identification
- Determine whether it is spam, two categories
- Various recommendations for music, merchandise, books, etc.
- Determine whether you like it or not, two categories
- Ad click rate estimate
- Determine whether you click or not click, two categories
- Voice assistants such as Xiaoai and Siri
- Identify your intentions, multi-classify
- Human identification
- Recognize whether it is a human face, two categories
In short, the process of computer learning knowledge from data has many similarities with the process of human learning, and it also has the following advantages: the speed and number of data acquired by machine learning is much higher than that of humans, that is, machine learning can be enabled by big data ; Machine learning can do one learning, multiple copy. Therefore, the name of machine learning is not called in vain.
If you like the above, welcome to follow me on my personal homepage: Homepage