The National Day holiday ended soon, and Gu Play returned to school with his sister and girlfriend and returned to normal study and life.

With the GPS and AMS projects coming to an end, Gu Yu has indeed been a little more idle in the scientific research tasks at hand.

Every day, as long as you go to class, take exams, bubble library, and even go to the laboratory. However, he is like a firefly in the night, destined not to let himself grow maggots.

So within a few days, he found Dean Ding.

"Dean, I want to send some interdisciplinary papers and engage in some interdisciplinary projects, this matter is not against the discipline of the school, right?" Gu Yu did not make an appointment, and walked straight into the dean's office at noon that day, and went straight to the showdown.

At that time, there was a vice president and the assistant to the dean. Dean Ding was slightly startled: "What's wrong? Are you wronged? Or do you suddenly have a brain twitch and want to change your major?" The vice dean next to him said in a round game: "Dean, don't worry too much, maybe it's because you can work hard and have a wide range of interests." Before Gu Wan answered, Dean Ding asked and answered with the vice dean, and made up for himself: "It turns out that you want to study the second major, with your current grades, of course this is no problem." You have been commended by the Ministry of Science and Technology, and the principal will give you special approval. What's more, our school originally allowed students to take two majors. "That's it?

I don't plan to double major in two majors, I just want to do a little bit of deep learning artificial intelligence across the circle!

But the school leaders have already decided so, and Gu Play will accept it. Everything was done quickly. It doesn't matter what the credits are marked on it, and if you have time to take a few more math courses, even if the standard of the second major is met.

I really don't want to fix it, as long as the paper is enough, it's the same. After getting this promise, Gu Play asked Ma Yiyi to try double cultivation to see if she could apply for it.

Ma Yiyi doesn't have such a big face, after all, she hasn't been registered at the commendation conference of the Ministry of Science and Technology, so everything still has to be done officially.

The result of the final consultation is that double majors are okay, but credits cannot be discounted and there is no discount. In the face of difficulties, Ma Yiyi couldn't help but ask a few more words about her husband's plans: "What do you want me to cultivate?" Gu Play replied very directly: "It's good to study cognitive neuroscience under the branch of psychology - this is a major where psychology and brain science intersect." Ma Yiyi looked confused.

Gu Yu knew that the direction he chose was almost the same as the path that Jeff Hinton took when he was engaged in deep learning on the earth.

Speaking from the heart, Jeff Hinton's abilities are not many anti-heaven scientists, but he succeeded. Eighty percent of this is hard work and the right direction, but at least two percent of it is luck and circumstance.

The success of deep-learning AI is due to the inherent technological biases of other earthlings – before that, few people had learned from it

From the perspective of "excavating the deep learning mechanism of the human brain", we will consider the problem of training machine learning.

So, whoever thinks so and is the first to put it into practice will be able to pick up the leak. Of course, this leak is not something that cats and dogs can pick up, and the argumentation process is more complicated.

In layman's terms, you must at least have a double Ph.D. in computer science and neurocognitive science from a world-renowned university, so that you are qualified to pick up this leak on the premise of being in the right direction.

A person who can't even do a double doctorate in an alliance school, the great opportunity is sent to him in vain, and he can't pick it up.

On the other hand, Gu Play is also well aware of another important opportunity for the development of artificial intelligence on the earth - before the development of deep learning and convolutional neural networks, the recognition of this technical route is not high, because even if this technical route is tested to 04 years, the preliminary principle is slightly run through, and the academic community still finds that there is a problem: the shortcoming of this algorithm is that the training efficiency is too low.

In other words, the computing power consumed is extremely large, and the progress of machine learning is actually very slow, and it is not possible to interpret white-box logic, so it cannot be explained through human intervention/

"Teach" to make the machine learn faster. There were two or three other technical routes at the time, and those routes had the opposite advantages and disadvantages of deep learning.

Those training methods are more white-boxed, i.e., the explainability of the decision-making process of machine learning, which is more understandable to humans, and therefore easier to intervene in, and convenient for humans

"Teach the machine to make rapid progress by hand". And the downside of these algorithms is:

"The upper limit of learning evolution is relatively low" (here is a general summary, but the actual scientific principles are much more complex than this.)

But to write to the layman, it can only be compared with this inaccurately. In other words, if humans continue to be bound by Moore's Law and rely on the technological improvement of computer CPU computing speed to meet the training computing power, then those AI learning paths that are more white-box and easier to initially train with a small amount of computing power may win in the historical choice

"Deep Learning/Convolutional Neural Network" is a technical route. However, in 2007~2009, the distributed architecture computing power design was born.

This thing first appeared in many Silicon Valley IT companies, and those programmers hated that the compilation speed was too slow after the code was written, and even felt that a dedicated compilation server alone was not enough, so they innovated one out of whimsy

"Distributed compilation architecture", so that all computers in the same LAN in a company can join this architecture, and then split the compilation task to all computers, so that all CPUs can jointly subcontract compilation.

Distributed compilation has been around since '07, and within a year or two, this thing has grown

"Cloud computing". After cloud computing, it is widely used, and the biggest impact of cloud computing on artificial intelligence training is that when an intelligence performs machine learning, it is no longer limited by the computing power of the CPU of the robot itself - the robot's own CPU is not fast enough, you can access the cloud, and use the CPU of tens of thousands of machines on the cloud to help you calculate together.

At this point, Moore's Law doesn't matter. As long as the computing power task can be split efficiently, a single CPU will be weaker if it is weaker.

As a result, the others

"Algorithms with higher computing efficiency, stronger white-box interpretability, but lower training ceilings and lower degree of automation" suddenly can't compete for deep learning.

Because the biggest bottleneck of deep learning was the black box, which was inefficient. But with the advent of cloud computing, computing power is instantly less valuable, and it can flood and train you.

…… Gu Yu knew these truths, but he couldn't tell Ma Yiyi everything. He could only encourage Ma Yiyi and let her lay out in this field, and the couple temporarily put together the peripheral technology first.

Fortunately, the timing is not very urgent. It took two years for Jeff Hinton to get the academic community to embrace deep learning.

As for the recognition from academia to industry, it took another four years - three of which were waiting for the emergence of cloud computing.

His achievements were initially recognized in the academic community in 06, but cloud computing officially appeared in 09, and Jeff Hinton was poached by Google with a high salary in 2010.

The artificial intelligence of later generations can summarize the principle of operation in one sentence, that is

"Using the computing power of cloud computing and using deep learning algorithms to process and learn big data". Computing power, algorithms, big data.

Among the three elements, in terms of technical difficulty, the first is big data, which is a resource with scarcity but no technical content, so it is the first to appear.

The algorithm, or the idea of deep learning, is the second to appear (06 years) The computing power to process these data with this algorithm, and the last one (09 years) After all three elements are gathered, Google on the earth began to run wild on this road the next year.

Gu Yu thinks that even if he opens the hanging, he will have more than 1 year to finish the architectural paper on the basic algorithm first.

The specific pursuit of industrial application will be talked about during the study abroad and company opening in the future. Moreover, the papers of the algorithm abyss stage were published when I was studying in China, and it is good to clear up the relationship with foreigners in the future - we did not come up with these things after studying in Stratholme, and at the beginning of studying at the Central University of Science and Technology, we have already made achievements in this area, and the credit is all Chinese.

After letting Ma Yiyi understand all these rhythms clearly, Gu Yu clearly assigned the task: "In addition to taking credits this year, I will give you two tasks to send papers." First, you are working on a distributed compilation job in the field of computer science. If you think of the architecture, whether it is a project company or a research institute of your Academy of Mathematical Sciences, try to do it first. If you really improve the efficiency, you can send out this experimental paper. Second, in the field of cognitive neuroscience, which is your double major, I hope you will post some research on the learning efficiency of human brain nerves - it has nothing to do with computers, that is, to study the learning efficiency of living people, do more control groups, and specific experimental designs, I will help you refer to them together. These papers will not directly generate commercial value, but they will make you and me qualified to be a little familiar with the top journals in related fields and enter the circle. "Gu Play has also published a lot of papers, but unfortunately they are all physics.

As the saying goes, interlacing is like a mountain, once he wants to enter mathematics/computer computing architecture/cognitive neuroscience, it is inevitable to brush his face and reputation again.

Just fight Warcraft with you, even if Orgrimmar worships it, the Undercity will have to be re-brushed. Of course, your Orgrimmar Worship will definitely help you farm your Undercity's reputation a little faster, and it's definitely easier than if you're a pure newcomer who doesn't have all the prestige.

Ma Yiyi still didn't understand the specific tasks assigned by her husband, but she felt that this did not prevent her from engaging in a personality cult and blindly following her husband's train of thought.

"I'll start with distributed compilation this semester. I'm a computer science student, and I think it's easy to write and publish. Ma Yiyi has also published papers in the past year or so in college, but only two in total, both of which are in the university's "Central University of Science and Technology Newspaper".

This result is already very good compared to other undergraduates, and even other top honor students who want to be exchange students.

But compared to her husband, it's simply not a dish. This time, she decided to drum up the distributed compilation architecture and send it to an external journal, preferably a foreign one.

() Sogou