Professor Cook's main research direction is "computational complexity theory", which most people must not understand.

Translated in two sentences, it looks something like this:

First, use a computer to hypothesize the human brain's processing model when faced with a problem. Pen Fun Pavilion wWw. ｂiquge。 For example, when a person walks into an auditorium full of people and wants to know the answer to the question "Is there anyone I know in the auditorium?" then he has to search step by step, from beginning to end, and if all the people don't know anyone after reading it, he can conclude that "there is no one I know in the auditorium". And as long as you find someone you know, you can make a judgment of "here are people I know".

Therefore, common sense generally holds that "from the perspective of scientific rigor, it takes more judgmental/computational resources to prove a thing than to falsify it, because falsification only requires finding a counterexample to end the argument and no longer continue to consume computational resources." And proof requires the reversal of all counterexamples".

However, in reality, the human brain is much faster than a computer when making a lot of rough judgments under the premise of "not deliberately pursuing absolute scientific rigor".

For example, if people look at a photo and judge whether the thing in the photo is a "cat", people can judge it at a glance, without verifying "whether the creature suspected of being a cat in the picture has biological characteristics such as XXXXX".

In other words, humans know how to grasp the big and let go of the small, and use the "fuzzy algorithm" to arrive at a barely usable but not very rigorous conclusion as soon as possible.

Before the 1980s, humans simply didn't know how to make computers "less rigorous".

Therefore, when computers solve all problems, they use rigorous brute force algorithms to carry them, resulting in many problems that cannot be exhausted because of the number of branching possibilities, and the computer cannot solve them.

For example, Go. Because even at the speed of computer hardware computing in the 2010s, if you want to exhaust all possibilities with "scientific rigor" of violent algorithms, the world's computers combined for distributed computing will not work. Therefore, under the guidance of that kind of thinking, human beings can only be satisfied with "using violent algorithms to overcome mental sports such as chess, which are exhaustive and incomputational". And the "alpha dog" in parallel time and space has killed so many masters, and it must not be close to the violent algorithm of imbeciles.

Stephen Cook's lifelong research is to solve the problem of "how to make computers learn to grasp the big and let go of the small like the human brain, and use limited computing resources to get a relatively accurate approximate result under the premise that resources do not allow it to be thorough and rigorous".

Gu Cheng feels that many computer science departments at the University of Toronto, under Cook's leadership, hide more talents who are trying to explore this field from different angles. And it is possible that Jeff Hinton just happened to be the first to attract attention in the field of artificial intelligence because of the choice of history.

But this does not mean that talents from other branches of the system are worthless.

If you can convince Professor Stephen Cook, it is obviously very helpful for Gu Cheng's overall and systematic poaching plan.

……

3 p.m., Mississauga Campus, Neural Networks Lab.

Gu Cheng saw Professor Jeff Hinton, who had already been cleaned up and looked cramped.

and Professor Stephen Cook, who has been famous for 20 years and has just arrived from St. George's main campus.

After a simple politeness, Gu Cheng said the main purpose of the trip, and first sent an invitation to Jeff Hinton, and the bid was very expensive.

"Professor Hinton, I can set up a research institute based on 'deep learning algorithms' for you, and you can personally get an annual salary of $2 million, as well as an annual research grant of $10 million, on a contract of at least five years. Your assistant and the graduate students with you, I can also give the best conditions. The only problem is that you could lose your place in academia for the rest of your life. Only a small percentage of your paper has the potential to be published, and you have to sign a non-disclosure agreement. See for yourself. ”

Jeff Hinton was a little embarrassed, after all, Stephen Cook, an academic who was 20 years older than him, was sitting next to him. Gu Cheng talked about money so bluntly, which was really insulting.

"We have never objected to doing things in enterprises, and academia should be combined with industry. However, I don't know what kind of topics a company that is obviously just doing social coercion and game comparison can make people come up with results that can be listed in the IEEE series of journals. ”

Professor Cook directly pointed out Gu Cheng's shortcomings.

Other industry giants, whether Microsoft or Google, have the support of top university research institutes. Although Gu Cheng is also engaged in the Internet, his technical content is the lowest type.

Just like the three giants of BAT in later generations, Tengyun is the one with the lowest technology content.

"You mentioned neural network algorithms as a proposed research direction, but I don't see how this relates to your industry. It is against the academic ethos of our university to have our professors at the University of Toronto do the kind of bells and whistles that don't see the academic future. ”

Professor Cook didn't care that he was just a director, and he just closed the coffin.

The so-called "neural network algorithm", Professor Cook has been involved in more than ten years ago, and compared with other "NP exhaustion theories" since the 80s, its biggest feature is that it "has no computing core".

Taking the human body as an example, as a biological individual, human beings have a central nervous system - the vast majority of body actions are controlled by the brain to control the limbs, and the eyes, mouth, ears, nose, hands and feet perceive external signals and then transmit them to the brain through the reflex arc (some of the lowest reflexes, at least until the medulla/spinal cord processing) and other brain processing instructions, and the hands and feet will respond.

But if the brain is dissected as an independent individual, the hundreds of billions of neurons inside the brain are equal. There is no such order that a small group of neurons is higher than other peripheral neurons, so that the information is preprocessed by that pinch of neurons, and then sent to the next pinch of neurons for processing.

(Of course, there are many other basic features of neural networks, but here we will only discuss the main differences from "genetic algorithms/annealing algorithms", so I won't go into details.) Otherwise, tens of thousands of words can be watered, and everyone still can't understand it. ）

When the concept of "neural network" was proposed, it was to explore a new way for computers to efficiently deal with problems similar to "find out if there is anyone I know in the auditorium": if there can be multiple computers, naturally randomly assign tasks, and use the nearest algorithm to find them from multiple points in parallel, then it is natural to speed up the solution of the problem by stacking up the number of CPUs when the performance of the "single-core CPU" is relatively weak.

But this concept does not "save computing resources", because theoretically it just turns "10 hours of work for 1 computer" into "1 hour of work for 10 computers". And these primitive "neural networks" still can't solve the "plausible" vague questions - they can only answer the either/or question of "there are people I know/no people I know".

Cook named Gu Cheng in this field and asked him to say that he was ugly about his application model, and Gu Cheng naturally couldn't avoid the war.

"I've read Professor Hinton's latest model hypothesis for neural networks, convolutional neural networks, and the learning algorithms that accompany them. I think this thing can be combined with the Internet's automatic identification/indexing tools. As for the specific application scenarios...... That's a trade secret, and I don't have anything to say. ”

"A new use of convolutional neural networks?"

Professor Stephen Cook was stunned, but quickly calmed down, he was not a concept that could be fooled.

"It seems that the focus of Mr. Gu's discussion is on 'convolution'?"

"Yes, if there is no 'convolution', only 'neural networks', we still can't talk about fuzzy problems that approximate the judgment of the human brain. Gu Cheng looked like he was in his chest, and he seemed to have predicted the other party's reaction.

He turned on his computer, plugged in the projector, and a picture of a cat appeared on the screen.

"I used the example of the cat in the picture - although the cat had one ear upright and one ear folded, the pupils of the eyes were a little abnormal, the tail was very short, and the coat color was dirty and very close to the background color of the photo, but as a human, I still recognized it at a glance that it was indeed a cat.

Now, I use the self-study program I wrote based on Professor Hinton's threshold idea to make a first prediction of whether the cat is a cat or not. In this algorithm, we preconstruct 30 combined feature quantities, such as 'cat eyes', 'cat ears', 'cat hair', and 'cat tail'...... Then, with the processing resources of 30 neuronal units, the results were predicted for each combination of feature quantities, and then the results were given separately.

Within these 30 neuron units, we make a judgment based on the pixel similarity between the cat's eye in this image and the cat's eye seen in this neuron', and give a product value to get a reference quantity such as 'there is an 85% probability that this is a cat's eye' or 'a 70% probability is a cat's ear'. Finally, the 30 combined features are combined according to the default weight of 1:1, and the final average score is higher than 60 points to determine that "this is a cat". ”

"That success rate must be pitiful. Professor Cook shrugged, compassion.

"Of course, it's pitiful, because my experiment has only just begun—to do this not to get the machine right, but to make the machine judge it again. If the judgment results of the machine and the human are consistent, then add 1 point to the current grouping of eigenquants and the weighted proportion array of each eigenquant.

Then, judge the next time. If it's still correct, add one more point. Until the judgment is wrong, then the existing weighted proportional array is automatically adjusted: for example, in the two previous 'average score of 60' results, the 'cat's eye' score is 75 and 80 points, respectively, and the 'cat's ear' score is 45 and 40 points. In the case of the incorrect judgment of the 'average score of 60 points', 'Cat's Eye' was 50 points and 'Cat's Ear' was 70 points. Then, we can conclude that among all the characteristic variables that determine whether a cat looks like a cat, 'cat's eyes' are the more critical variable than 'cat's ears', and their weight should be increased when calculating the composite score.

Finally, according to this logic, let this algorithm look at a hundred cat pictures, a thousand, ten thousand...... The algorithm will naturally summarize a set of judgment weights that are 'not all right, but the probability of being correct is getting higher and higher'. ”

Human children, at the age of 3, learn to recognize all kinds of things, in fact, this is how the brain calculates. There are no characteristics that must be insisted on, look at hundreds of cats, and naturally adjust the weights of each feature to know what a cat is.

There is no variable that has the power of "one-vote veto". At best, it is just that its "integral" in convolutional neural networks is relatively high. It is precisely because of this that humans can recognize a cat when they see a cat whose eyes have been completely gouged out.

……

Gu Cheng's overall exposition is naturally very lengthy and difficult to repeat one by one.

Many of these tricks are completely easy to understand after being explained thoroughly, and there is nothing compelling at all.

However, Gu Cheng at least provides a path for convolutional neural networks that "if you can't do all the right, there is no commercial value", and "even if you don't do it well now, you can still achieve the possibility of phased commercial realization within one or two years".

Professor Stephen Cook talked with him for a long time, but finally did not know what to do.

"The sacred research on neural network algorithms is explained by such unrigorous and unscientific inferences, assumptions, and simulations. These ideas and inferences simply cannot form a system of papers and results. Professor Cook instinctively lashed out a few words, but after calming down, he gritted his teeth and had to admit, "But, very enlightening." ”

Gu Cheng was unimpressed by Professor Cook's accusations: "It's like Chinese medicine, it's not scientific, but sometimes it can really cure diseases." It's just a matter of luck, and I can't explain the inevitable reason why it cures the disease -- but I only want the curative effect, and I don't care if the science is unscientific or not. You are interested in proving and perfecting the academic system, and I respect your ideas. But I'm not going to pay for these proofs. I just want to invest in something that can be used even if it's not scientific. ”

Professor Cook shook his head helplessly: "It's a pity that I have such a good brain, but I am not proud to devote myself to science." ”

"No way, I'm an industrialist, and what I care about is pragmatism. ”

Gu Cheng no longer paid attention to the idealistic old scholar, and just re-proposed his invitation.

"That's all I have to say today, Professor Cook. I hope you don't use your academic authority to dissuade others from joining industry. I would also like to welcome you to introduce me to talents who can 'try their luck' from various angles. ”

On the East Coast in 03, there are still a lot of intertwined mixed scholars, and the combination of achievements and industry is generally still stuck in those achievements that can brush both face and money.

It wasn't until Jeff Hinton figured it out that the East Coast academic community was going to have a strong atmosphere of measuring success by money. Everything comes step by step.

"If any of you want to go, go. I'm not going to stand in the way. As for the second suggestion, I will consider it. ”