Joint compilation: Zhang Min, Gao Fei
When I participated in computer science in 1995, data science did not exist, but we are still using many existing algorithms. This is not just because of the return of neural networks, but it may have been that since then there have been no fundamental changes. At least it feels like this to me. Interestingly, from this year onwards, artificial intelligence seems to have finally become mainstream.
Before the advent of cloud computers, smart phones, or chat bots, 1995 sounded like a very painful period. But in understanding the past few years, if you are not in the environment, it feels like it was a long time ago. Like self-continuity, it sticks everything together. Although it has changed a lot, compared to the present, the world does not feel fundamentally different.
Perseverance in computer science
Computer science has never been as close to the mainstream as it is now. Later, with the first cyber bubble around 2000. Some people even questioned my choice of learning computer science, because it was very easy to program computers and no longer needed an expert.
In fact, artificial intelligence is one of the main reasons why I study computer science. It is intriguing to me to use it as a constructive way to understand human thoughts. I trained in the first two years to ensure that I had enough mathematics to solve the obstacles on the road, and eventually I went to the first AI class (teaching by Joachim Buhmann), when Professor Sebastian Thrun of the University of Bonn was preparing to leave United States. I have to check which of my speech periods I participated in. He has two lectures in computer vision. One is pattern recognition (mostly the old version of Duda & Hart's book knowledge) and the other is information theory (similar to cover and Thomas book of). The material is very interesting, but it is also a bit disappointing. As I now know, people stopped the symbolic work of AI and no longer insisted on using more statistical methods. The essence of this method of learning is to reduce the problem of selecting the correct function based on a limited number of observations. Computer vision lectures, even learn less, and rely more on explicit physical modeling to get the correct estimate, for example, to reconstruct the motion from the video. The method at that time was more physiological and physical than it is now. Although neural networks exist, it is clear to everyone that they are "another function approaching." Everyone thinks this way except Rolf Eckmiller, and another professor whom I have worked under. Eckmiller built his entire laboratory on the premise that "neural computing" is somewhat better than traditional calculations. This can be traced back to the days when NIPS had a complete path dedicated to the study of the physiology and working mechanisms of neurons. Some people even thought that there are essential differences in our brains that may occur at the quantum level, which increases the human spirit. This difference is a major obstacle to the study of truly intelligent machines.
Although Eckmiller is very good at promoting his opinions, most of his staff are glad to be down to earth. Perhaps this is a very German thing, but everyone is very concerned about whether these computational models can be achieved. This is also a problem that has been plagued by me in the research. I graduated from October 2000 and published a fairly far-fetched master's thesis, trying to establish a connection between learning and working on optimization issues. Then I started my doctoral dissertation and stayed in the field until 2015.
There are many research methods for machine learning, but the basic problems are basically the same
Although it has been trying to prove the relevance of the industry, it is a very academic effort for a long time and the community is quite closed. There are some personally successful stories, such as handwritten character recognition, but many companies fail in machine learning research. I remember a company called Beowulf Labs and NIPS who used a video to recruit talent everywhere and promised to be the next “mathtopiaâ€. In essence, this is the story of DeepMind, recruiting a group of outstanding researchers and hopefully it will take off.
The entire society will also revolve around one fashion to the next. One strange thing is that machine learning as a whole, in addition to a lot of methods, there are only a very few fundamentally different issues, such as classification, regression, clustering and so on. It is not like physics (I suppose) or mathematics, some difficult problems that are generally considered unresolved, and solutions that can advance the best results. This means that progress is often done horizontally, and by replacing the existing method with a new one, the same problem is still solved in different ways. For example, first there is a neural network. Then a support vector machine emerged, claiming that it works better because the relevant optimization problems are convex. Then there are boosting, random forests, and so on, until they return to the neural network. I remember that the Chinese Restaurant Processes “fired†for two years, but nobody knew what they meant now.
Big Data and Data Science
The era of big data and data science is already here. At that time, I was always in the academic world. I always felt that big data and data science must have come from the outside world. It may have come from companies like Google that handle huge amounts of data. Large-scale learning does exist, for example, genomic data in bioinformatics, but we should look for more effective or approximate algorithms to solve these problems rather than be reckless.
Companies such as Google have finally confirmed that we can use huge amounts of data to accomplish some things that will eventually change the mainstream view that people hold about artificial intelligence. Some technologies, such as Hadoop and NoSQL, appear to be very popular and can be skillfully promoted and continuously innovated in the market. These technologies will be free from technical limitations in existing systems. However, what effect will this have on machine learning researchers? The impression given to me is that these researchers are pleased that the arrival of the age of big data and data science led them to finally gain recognition. However, they also feel sad about the way they get their approval. To understand this, we must realize that most ML researchers are not scientists in the computer field, or that they are very good at coding. They are very interested in coding. Many of them are majoring in physics, mathematics or other disciplines. In these disciplines, their excellent mathematical training ability enables them to skillfully use various algorithms and build core models for machine learning.
On the other hand, the Hadoop distributed system infrastructure is more technical. Written in Java, this language was considered highly professional at the time, and compared to MatLab and Python, the language of Hadoop is incomprehensible. Even those C++ programmers will feel the same way, and for them, the Java language will be verbose, especially in numerical computation and simulation. However, currently there is no way to solve this problem. Therefore, they have renamed everything they have done as big data, or they have begun to emphasize that big data only provides basic data resources for large-scale computing, and you need professional talents. Can understand this data information.
This solution should be no mistake. In my opinion, to a certain extent, this distinction still exists. If you choose the right language for data analysis, Python is the best choice. There are some technologies, such as Spark, that try to analyze data by binding to the Python language, regardless of whether this method makes sense in terms of performance.
The return of deep learning
Even before Google developed DeepDream, an artificial intelligence technology, neural networks have set off a return boom. Some people, such as Yann LeCun, have always insisted on this method. However, some 10 years ago, some studies showed how to use stratification training and other training methods to train “deep†networks. The size of this type of network is beyond. What people can imagine before.
The evaluation is based on training examples to train the neural network, and then adjust the ownership value to further reduce the error. If the gradient value is recorded in the direction of decreasing weight, the error will be propagated back from the last layer. In any case, it can be understood that the error information will be reduced by layers, which will increase the difficulty of multi-level training networks. As far as I know, many people still use the background method. I am not sure whether this view is still correct. However, it is certain that the amount of data that can be used, the tools, and the original computing power have all changed. As a result, some of the first research literature ignited people's interest in neural networks. People then began to use these neural networks and achieved remarkable results in some fields of application. These neural networks were first successfully applied to computer vision, followed by speech processing. Processing and other fields.
I think that this type of neural network can attract people because of its diverse uses. With this approach, people can avoid the hassle of understanding a variety of different approaches. In addition, the neural network has a modular structure, and people can pick out and combine different levels and structures to solve various problems.
Google had published an excellent paper on DeepDream artificial intelligence technology. It was mentioned that they could use some information-rich network to generate some data. However, we humans have the ability to learn structure and attributes on the fly and learn to apply it quickly. This kind of network. Thus, Google can now be regarded as a first-rate artificial intelligence company. Artificial intelligence will save the world.
A basic problem that remains to be solved
I have communicated with many scholars who are dissatisfied with the leading role of deep learning because this method can produce good results. Sometimes this effect is even too idealistic, but it still cannot help us to further understand the human brain. working principle.
I also hold the same view on this, that is, this basic issue has not yet been resolved. How do we understand the world? How do we create new concepts? Deep learning still stays at the level of imitative behavior, although for some people, the effects of deep learning have been quite good, but I'm not quite satisfied with it. In addition, I think there are risks to attribute too much intelligence to these systems. For the raw data, the performance of these systems may be very good, but in other aspects, these systems will operate in completely different ways.
Although Google's translation tool allows people to skip content posted on foreign websites, the performance of such a system still needs improvement, which is unmistakable. Sometimes I don’t think anyone would care about this, maybe it’s because no one will be harmed, right? However, it may also be due to my German cultural background. I hope we can see the development of these things in the first appearance of things.
Via: original
PS : This article was compiled by Lei Feng Network (search “Lei Feng Network†public number) and it was compiled without permission.
Second-Hand Power Generator,Power Generator,Diesel Generator,Electric Generator
Shaoxing AnFu Energy Equipment Co.Ltd , https://www.sxanfu.com