Ian, who was born in '85, has been a bit stressed in recent years.

For an ordinary doctoral student, there is nothing wrong with him, on the contrary, he is smart and hardworking, and he has been studying in a famous family.

He graduated from Stanford with a bachelor's degree and a master's degree, and his mentor is Andrew Ng, the founder of Google Brain, and his Ph.D. study in Montreal under the guidance of Bengio, one of the three giants of AI.

It is not an exaggeration to say that his academic experience and personal wisdom are one in a million, but if you want to compare it with his mentors, it is indeed a little bleak.

The 28-year-old Ian is in his final year of Ph.D., and data generation algorithms are the direction he has been thinking about since his Ph.D. years.

Many researchers are already experimenting with neural networks to simulate the human brain to generate new data.

Unfortunately, so far, this data is usually so problematic that it is completely unusable.

And a big obstacle is that the resulting images are often indescribable. These useless images lack the means to make use of them, and they are not easy to evaluate.

A FEW DAYS AGO, THE RESULTS OF THE MODEL PERFORMANCE ON IMAGENET THAT EXCEEDED HUMAN STANDARDS GAVE HIM A NEW INSPIRATION.

If deep networks can already reach the human level in some image-based tasks, can we avoid the problem of poor results of generative networks and complete such a process through the confrontation of two networks?

The more Ian thought about it, the more reasonable it became, and instead of even going to sleep last night, he spent the night implementing a prototype version of the generation-confrontation, and tried to train it, and found that the results seemed to be basically the same as expected.

Unable to contain his excitement, he wanted to discuss the principles and practices with his mentor as soon as possible.

Ian wandered back and forth in front of Bengio's office for more than twenty minutes before Bengio strode over from a distance.

If you ignore the white mixed in his hair and beard, the 49-year-old Bengio actually looks quite young, and in terms of this face, he is at most a man of thirty-five or sixteen.

In particular, the two thick eyebrows, which are shaped a little like Mr. Bean, make him more loyal and approachable.

Ian hurriedly took a few big steps forward, unfolding the notes and experiment notes in his hand at the same time. He pointed to the core of the above lines, and was about to speak, but was ruthlessly interrupted by Bengio.

"I know you're in a hurry, but don't worry about it yet." Bengio looked at Ian's anxious appearance and joked, "What we pay attention to in doing academics is to have a solid foundation and work steadily. What's the use of being in a hurry? It's useless to be in a hurry. ”

"What's the other thing you can't say when we go in and sit down and talk slowly?" Bengio said, pulling out his key and opening the door to his office. "Come in, it's only past seven o'clock and you're crouching in front of my office, I think you must have made some major discovery, right?"

"I'm in a hurry, I'm really in a hurry." Ian had been working in the generative space for a while, and by the time he thought about it yesterday, he knew it was going to be a major change in the paradigm.

From the beginning of his undergraduate studies to now, ten years of sharpening his sword, Ian is very excited. He has a good academic sense and can sense the extraordinary nature of the idea. Is the day of your own emergence approaching?

At this moment, he could feel that this was an excellent opportunity to use his own strength to get rid of the experience of being a disciple of Wu Enda and a disciple of Bengio, who was both a halo and a shadow.

Let the world truly recognize himself, and re-see himself through the techniques and algorithms he proposes.

"Is the two networks pitting against each other?" Bengio pondered for a while, not in the rush to raise objections and skepticism like Ian's friends, Bengio has always been very encouraging of imaginative, creative ideas.

"Train two models at the same time, not one. The generator and the discriminator can be stepped on with the left foot and the right foot spiraled to lift each other. You want to use this pattern to make the distribution of the generator as close as possible to the real data provided. ”

"In this way, we also get rid of the usual way of thinking, and we train the discriminator not so that it can tell 100% whether an image is generated or not. Rather, our goal is to make it clear to the discriminator what is true and what is false. Ian added.

"Adversarial is just a formality, and its purpose is to train a generative network. I'm confident that this approach can get rid of the dependency on data labels, and a lot of other data can be leveraged without additional processing. ”

Ian's agitation had not subsided until this point.

"More than that, generation is just the most preliminary application." Bengio thought for a moment and said, "It can be used for a special purpose by sifting through the data. With only a small amount of code tweaking, I believe it works well with simple layouts, especially in directions like faces, which is a very dangerous technique. ”

"What are you going to give it a name?" Bengio turned on the computer as he spoke, naming technology is a difficult problem and an art.

If you don't choose a good name, your work will be done in vain.

A loud name, a funny and concise abbreviation, all have a good chance of boosting the visibility, popularity, and application of the work.

"Generate... And.. Antagonism. Or call it a generative adversarial network? Without even thinking about it, Ian added two of the most crucial elements of the idea to the name.

"Generative Adversarial, I'll take a look." The first step in choosing a name is naturally to search to see if anyone has already occupied the pit.

Bengio's retrieval, Meng Fanqi, who was far away in the sky, suddenly sneezed, and said in his heart that in this big summer, why did he still catch a cold?

"It seems that this name has been taken up by someone, just two days ago, which is very unfortunate." Bengio didn't realize the seriousness of the problem at this time, "It looks like we're going to have to change another name." ”

As he spoke, Bengio clicked on the preprint.

"What kind of network is DreamNet, I don't think I've seen it before." Bengio vaguely felt that the name seemed to have been mentioned by someone in the past two days, but he was busy with other research matters and didn't pay much attention to it.

But as soon as he finished reading the summary, Bengio's eyes sharpened and his expression became very serious.

Ian was still pondering what a name would be better.

Bengio, on the other hand, ruthlessly pierced his defenses with one sentence, "Ian, it looks like we can't use more than just a name." ”

"What? Our thoughts crashed? "Ian was shocked, and his academic ideas collided, which was probably one of the most unfortunate things for researchers.