Take myself as an example. I work in cv direction (computer vision), and parameter adjustment doesn't take a big part in my daily work.

In CV, besides superparameter, the main factors affecting the model effect are network structure, data and loss function. After these three aspects are determined, there is not much time to adjust the parameters.

Back to the point!

Turn it in. Turn what? In artificial intelligence, parameters can be roughly divided into two categories:

Why is it called "the person who adjusts the parameters"? There are two main reasons:

How to avoid becoming a "trainer" At present, the competition of AI talents is becoming more and more fierce, and the era of "trainer" has slowly passed. In fact, these things don't need AI engineers at all, and future R&D engineers can undertake them! A few years ago, if you were proficient in TensorFlow and mastered the basic AI algorithm, you could easily find a high-paying job. But it's different now. The requirements for AI positions are getting higher and higher, and the depth of knowledge is also higher.

If you want to keep up with the times, you have to arm yourself so as not to be eliminated.

For real artificial intelligence engineers, they often start with data and characteristics, and also need rich industry experience. Be sure to remember an industry proverb, data and features determine the upper limit of the algorithm, and the selected algorithm and parameters only determine the speed of approaching this upper limit.

There is no shame in adjusting parameters. Okay, the parameter adjustment man is very good. Technically, algorithm engineer only deals with data and models. The model is a black box, which comes from data and parameters.

There are two parameters in the model, one is called weight, which can be learned; One is called superparameter, which needs constant experiments to determine. The so-called tuning is the latter of tuning. Of course, these experiments need professional design skills, which are beyond the scope of this article. Interested students can look for Andrew Ng's books.

Many people say that algorithm engineer is a detective and has no technical content. They are all xgb. Why can some people win the championship and some people can only be weak? Maybe you will say that feature engineering is well done. But in the field of graphics and text, the model is based on building blocks, which can also be regarded as superparameter, the number of layers of the model and the dimension of the model.

In a super-resolution competition, a team from South Korea won the championship, taking away the batch standardization that everyone took for granted, and unexpectedly won the championship.

Practice is very important, and there is no shame in adjusting parameters. After training, you can win the championship. You can even write a paper on the experience of parameter adjustment. At that time, Google had a paper that violently tried various functions and published a paper.

Sometimes the theory is to practice first, and then guess or infer. Right? Never mind black cats and white cats. A cat that can catch mice is a good cat. Goal orientation.

Not all Andhadhun can surpass the wonderful music. Although there are only a few notes. If the skills of parameter adjustment are well done, papers can be published, business indicators can be improved and profits can be brought.

There may be two paths: from theory to practice, or from practice to theory. Unfortunately, most people can't do it.

It's not that easy to be a good detective. The key is thinking. People who are good at thinking and reflecting, whether in theory or practice, will make faster progress than mechanical repetition and become heroes more easily.

In fact, the meaning of "Diao Shenxia" is similar to that of "code farmer" who writes programs, and it is a mockery of people engaged in this industry. For example, if you write a program, you will be exposed to the business of adding, deleting and checking at first, and if you do more, you will say curd. There are many other names for artificial intelligence, such as Bao Xia, an indicator slave.

The technology and knowledge of artificial intelligence are still very extensive, not just tuning parameters. There are also data and feature engineering, mathematical algorithm knowledge and so on.

Twenty years of professional brick moving

this ...

I saw this name for the first time, but this problem does exist in the field of artificial intelligence.

To give a simple example, let's take the hot artificial intelligence method: deep neural network.

So what is a deep neural network? Let's talk about neural networks first. As the name implies, neural network is a theoretical algorithm for scientists to realize artificial intelligence by simulating the coordination of human neurons.

Let me explain here that the so-called neurons are the bearing part of human thinking activities, and some thinking activities in the brain need the participation of neurons.

So what is a deep neural network? To put it bluntly, it is the superposition of several layers of neural networks. More specific principles, I won't go into details here, this simple statement of deep neural network, we can give a more vivid example to illustrate:

Now we need to solve a problem, that is, how to identify the animals in the photos, whether they are people or some other animals. We now use the method of deep neural network to identify this picture.

So what is a person's distinguishing feature?

The most common ones are head, trunk and limbs. But there is a problem here. Just like a common kitten in daily life, a puppy has a head, a trunk and limbs. Therefore, the problem of 1 layer of our deep neural network is to distinguish the outline of this person from the outline of cats and dogs. Here we want to introduce a concept of contour similarity, which is a simple topological meaning. I won't go into details here. If you are interested, you can search the related content of topology yourself. When the 1 layer of this deep neural network is used to distinguish people from cats and dogs, there will be a parameter problem of the neural network, and we need to adjust this parameter to an appropriate degree to distinguish people from animals.

Of course, with the 1 layer, we may have a second layer or more, such as whether this person is wearing clothes or not, and the skin color of this person, such as the hair of this person growing on the head, not on the whole body. At this time, this layer of neural network will involve a parameter adjustment problem. Only by adjusting the parameters to appropriate values can the machine correctly identify whether to wear clothes, whether to have hair on the head, and so on.

In short, we can see from the process of deep neural network identification that these artificial intelligences are actually the process of adjusting some parameters in the existing model, so the parameter adjustment man in the landlord problem is really worthy of the name.

It is too difficult for a few people to come up with new models and new ideas. They are all alchemists, and it is difficult to explain why.

People who do artificial intelligence basically deal with data models. In addition to data, the data model also has model parameters, which are adjustable. What we usually call hyperparameters is to adjust them to the data. Of course, adjusting parameters also requires mathematical skills and understanding of algorithms. So it is impolite to simply say that people are adjusting parameters. You may not be able to do what others can do.

There are few effective models, and the remaining engineers really just need to adjust the parameters and find data to train! [Laughter]

People who can only use other people's models are actually a misunderstanding and prejudice of the outside world and those who are superficial about artificial intelligence. Of course, artificial intelligence can't just use other people's models.

Because after the algorithm is encapsulated, the only thing left on it is the tuning parameters. Data cleaning, feature selection and feature engineering are not technical activities, and there are many capable people. You need to know the algorithm to adjust the parameters.