Easily commented: Wu Enda ran away, Baidu's genes were not acceptable?

This article is the "Easy Review" section of Netease Technology

Wen/Binghan

Wu Enda is gone, Baidu is silly. Baidu, which is preparing for quick-drying artificial intelligence, suddenly makes people feel at ease. Wu Enda’s open letter does not explain anything. After all, we haven’t seen open letters of resignation. When Wu Enda was at the time, what Baidu might face would still be a problem of realizing cash. When Wu Enda left, it became a problem of technological evolution. Wu Enda spent only three years in Baidu. He brought to Baidu not only several wonderful papers but also various commercial artificial intelligence products. Although these products have encountered the realization of the problem, but also let Baidu in the field of artificial intelligence in the front.

I don’t dare to discuss why Wu Enda left, but Baidu’s field of artificial intelligence is not without taint. Even under the leadership of Andrew Ng, Baidu’s character has not changed. In 2015, Baidu achieved the best global results in ImageNet's image recognition artificial intelligence benchmark. Its image recognition error rate was only 4.58%, while Microsoft was 4.94% and Google was 4.8%. But then ImageNet issued a statement stating that Baidu had irregularities in the test: Baidu researchers used intensive testing to improve their performance. Between November 28, 2014 and May 13, 2015, Baidu's heterogeneous computing team researchers used at least 30 accounts to submit at least 200 test server visits, including five from March 15 to March 19. More than 40 submissions were made during the day, in violation of the maximum of 2 submissions per week. As a professor who has fought for many years in the academic world, Wu Enda sees this situation as an estimate that she should be on her own. If we do such a rotten thing as a student, we need not say that graduation is hopeless and the possibility of being expelled is also great. However, as a Baidu employee, the statement of apology was all gone.

Wu Enda's deep neural network has brought surprise to Baidu artificial intelligence, but it has not been able to pull Baidu from the quagmire of decline in search portal share. Companies such as Baidu, and today’s headlines, may not be helpful to their commercial success. With tens of thousands of GPUs, Wu Enda may bring some surprising artificial intelligence products to Baidu, but it is difficult to reproduce the profitability of the Putian Department of Printing.

However, I've been too lazy to spit on Baidu or major Internet companies to face the "neuropathy" behavior of artificial intelligence. This article is just to prepare the nerve network technology that makes artificial intelligence stunning the world to commemorate Wu Enda's outstanding contribution to the field of artificial intelligence in China, especially in the field of neural networks.

Why is neural network fire?

Neural network This thing is not new, AlphaGo cattle X is just a branch of various artificial intelligence algorithms. Many algorithms in neural networks have existed for many years, and it is not surprising that they suddenly appear like magic. Recently it has begun to attract focus, mainly because of the rapid development of available computing power (CPU, GPU, AI-specific computing units), which makes mass matrix multiplication operations easier to test, verify, and iterate.

However, Deepmind's optimization of the neural network of the AI ​​game under the AI ​​is truly amazing, and Wu Enda’s project to identify cats on Google is also bringing new levels of artificial intelligence to image recognition. As a result, the people who do not know the truth really think that artificial intelligence will come to replace humans tomorrow. Major Internet companies also followed suit and immediately started an artificial intelligence business.

However, even if Wu Enda Google Baidu has established a neural network with a number of billions of links, it is still somewhat early for artificial intelligence to replace humans. The acquisition of training data in different fields is still difficult to solve. And in some areas of data complexity, perhaps only quantum computing can end training within the acceptable range of human life.

The neural network or the key to the Turing but it does not solve everything

Neural networks have two major benefits. One is the use of relatively long pre-application training time in exchange for real-time processing in applications. The second is to use the processing complexity of training data in exchange for the processing complexity of the preset logic. And AlphaGo made people believe that maybe neural networks are really the key to solving the complete Turing test.

In terms of macro desires, neural networks are expected to directly realize artificial thinking by simulating realistic biological thinking. An artificial neural network is an adaptive machine that models the human brain to complete a specific task or method of interest. It is a large scale parallel distributed processor composed of simple processing elements. Naturally has storage experience and features that make it available. Neural networks are similar to humans in two aspects: 1. The knowledge gained by neural networks is learned from the external environment; 2. The strength of the connected neurons, synaptic weights, is used to store acquired knowledge.

There are many problems in real life and cannot be solved with preset logic. When the problem is unpredictable, such as in handwriting recognition; or the problem of processing can easily change the requirements; the task of processing needs only a satisfactory solution instead of an exact solution. These conditions are not a problem for the fuzzy logic processing of the human brain. However, it is a problem for computer programs. The neural network can complete an optimal solution by having a large amount of empirical data that can be referenced, or the task itself can generate enough empirical data. It does not guarantee 100% correct task completion, but it can give an optimal solution to the experience like the human brain. From the perspective of mimicking human thinking, artificial intelligence may be the most versatile of all machine learning algorithms, and it is also the most likely approach to touchdown testing in touchdowns (editor's note: rugby terminology, completion scoring).

But what does this mean that neural networks will solve everything? It is nerves. In many cases, neural networks are not the optimal solution for artificial intelligence. What issues are taken up by neural networks? That is neuroticism. Here is an example - basic digital identification.

Whether it is a neural network or a basic decision tree, or a vector machine, the most fundamental algorithm selection principle can still solve the problem. Many products are processed using popular recursive neural network algorithms, but if you can use simple preset logic, Bayesian algorithms and vector machines can achieve good results, there is no need to put the cart before the end. Deep neural network is likely to become the universal algorithm to solve the artificial intelligence problem in the future, but the large amount of data preparation and finishing that the neural network needs is also time-consuming and time-consuming. Compared with other machine learning algorithms, the neural network is more like using the large amount of data clean-up time and training time before the application for rapid calculation in use.

Comparing the performance of various algorithms, the neural network is not the optimal solution.

Source: (artificial intelligence, Stuart-Russel)

At present, no artificial intelligence algorithm can achieve the optimal solution in any scene. The fact that neural networks are fired does not mean that artificial intelligence in the future will go completely in a single direction. Neural networks combined with other methods of assistance may be a more sensible choice.

Artificial Intelligence Realization Challenge

Wu Enda was recruited by Baidu in 2014 and Baidu started the age of artificial intelligence. You can say that Baidu has been imitating Google, but in artificial intelligence, Baidu has really dug up the core of Google's brain. Wu Enda published nearly 20 papers in Baidu's 3 years. Each of the results was either integrated into Baidu's existing product line or based on the results of a start-up company. From Wu Enda back to Baidu's first medical inquiry robot developed to the second phase of wake-up product of the recently launched speech recognition development platform, Baidu's progress in the field of artificial intelligence is obvious to all. Although there is no such thing as Deepmind, it has become an industry leader.

Nowadays, with AlphaGo's east wind, various artificial intelligence companies have mushroomed. Not only have all kinds of big data companies transformed themselves into pioneers of AI, even media companies that rely on data mining success have claimed that they are an artificial intelligence company. Looking at the artificial intelligence from the outside is like a treasure island. Tens of thousands of people fought over the single-plank bridge to dig out only the treasures on the island. However, the biggest problem that Yokohama faces in all artificial intelligence companies is how to realize it. Returning to the actual level, there is neither gold nor silver on this island. It does not look like a decent piece of stone.

Taking Baidu as an example, Wu Enda is an authority in the field of artificial intelligence picture recognition. However, Comrade Enda has never had a successful case of commercializing products. When Stanford taught as a professor, he was supported by the school. Of course, Google's brain was kept by advertisers. There was never any pressure for liquidation. After coming to Baidu, of course, the pace of productization has accelerated, but it can only be the rhythm of being raised. For example, Baidu's inquiry robots, which were launched in 2015, are not a problem in promotion. The logic of relying on Baidu's inquiry and selling drugs should have a good prospect, but last year's Wei Zexi incident. Basically, Baidu completely crushed the logic of the realization of medical violence. As a result, the medical division became Baidu medical brains, and the inquiry robot products could only become tasteless. Of course, some of Baidu’s voice products still have a very good toB outlook on the technical level, but this prospect is not likely to become a moneyscape compared to Baidu’s mass and artificial intelligence.

Wu Enda's departure is a heavy blow to Baidu, and Baidu wants to find another person whose level is similar to that of Wu Enda. It should be said that this is basically impossible to achieve. Li Yanhong's wild wilderness facing Bell may be prudent, but in the face of his own AI, he will inevitably mourn. Someone previously said that Wu Enda was sent by Wu Chengen in order to help Baidu invite more Sun Wukong. If more Sun Wukong joins, the story of Wu Enda's "Artificial Intelligence Changing the World" may be even more exciting. But Wu Enda is gone. Is it true that the Sun Wukong will also leave. Baidu artificial intelligence may be faced with a crisis for us to wait and see.

Editor's note: The author of this article is Liubing Han, CEO of EditorAI, an artificial intelligence startup. The author writes the most articles in the field of artificial intelligence, and writes articles that have the most knowledge of artificial intelligence.

Magsafe Power bank

magsafe power banks ,magnetic power bank ,wireless power bank for iphone ,apple magsafe power bank ,magnetic wireless power bank

Pogo Technology International Ltd , https://www.wisesir.net

This entry was posted in on