By Super Neuro
Regarding machine learning, all of you veterans should be familiar with it.Among the many explanations of the concept of machine learning, Cassie Kozyrkov, chief decision intelligence engineer at Google, called it a "thing labeling machine," which is an interesting understanding.
Machine learning is essentially a thing labeling machine
In addition to being Google's chief decision intelligence engineer, the beautiful Cassie Kozyrkov is also a statistician and neuroscientist.
Her understanding of machine learning is different from the general mainstream view. She believes that machine learning may not be as magical as experts say, nor is it the main reason for attracting $30 billion in venture capital to the field of AI, nor is it as profound as Hacker News or Zhihu say.
In the eyes of Miss Cassie Kozyrkov, the above are all written explanations of machine learning. In actual application scenarios, machine learning is essentially a "thing labeling machine." By recording your description and labeling it with a corresponding label, it guides the computer's next action.
Machine learning is more practical than science fiction
Machine learning is a subset of AI and one of the key technologies for realizing AI. Currently, most AI products on the market basically rely on machine learning, so many people regard these two concepts as the same.
Science fiction imagination about AI has also unconsciously shifted to machine learning. For example, Jarvis, Iron Man’s AI assistant in the movie Iron Man, can always instantly locate criminals on unknown streets in a remote country. In reality, it is difficult for machine learning to achieve the scenes in science fiction movies, and even current AI cannot achieve them.
At present, whether it is AI or machine learning, they are widely used to improve computer efficiency and expand application scenarios. They can be used to process huge data projects and solve some procedural tasks. Compared with science fiction films, they focus more on practicality.
For example
Below is a picture of a cat, which the human brain can easily recognize through various senses and experience, but this requires a lot of "mental activities" for a computer.
Given a computer a task: classify (or label) photos as cat/not-cat? Machine learning systems and traditional programming methods will give you two different operating experiences.
In the traditional programming approach, human programmers think carefully about pixels and labels, communicate with others, inspire ideas, and finally manually build models.
A model is a set of instructions that a computer must execute to convert pixel data into labels for the computer to recognize. These instructions are just some code that the computer uses to convert input into output, and they can be written by a programmer or derived from data through an algorithm.
Let's take a more complicated example.
How to use code to describe each pixel block contained in this picture?
This is very difficult for the human brain. The human brain can recognize the photo, but it is difficult to encode the pixel blocks in the photo. Therefore, the human brain giving instructions to the computer to recognize the photo is not only a lot of work, but also very complicated and impractical.
Therefore, traditional programming methods are difficult to apply in the field of image recognition.
But machine learning can solve this problem very well. It is a completely different programming paradigm. It can be programmed through a classification-like technique without the need for explicit instructions.The official explanation is: find fixed patterns in the data and convert them into instructions.
Taking the above picture as an example, machine learning will integrate all relevant data, summarize a bunch of "not cat" examples, and a bunch of "cat" examples, and then reclassify them according to relevant features until "cat/not cat" is determined.
Machine learning can express many ineffable things in computer language, which means that we don’t need to give specific instructions for the computer to get the results we want.
This is also the main purpose for which AI and machine learning were created, to interpret human intentions without the need for instructions.
This transformation has made computers increasingly intelligent, able to solve many problems that only the human brain can solve. This is a qualitative leap in human technology and a sign that computer science is knocking on a new door.