Ian, who was born in 1985, has been under a lot of pressure in recent years.
For an ordinary doctoral student, there is nothing wrong with him. On the contrary, he is smart, hard-working, and has been studying in a prestigious school.
He graduated from Stanford with both undergraduate and master's degrees. His mentor was Andrew Ng, the founder of Google Brain. He studied in Montreal for his PhD and studied 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 compared with his mentors, they are indeed a little bleak.
Ian, 28, is in the last year of his Ph.D., and data generation algorithms are what he has been thinking about during his Ph.D.
Many researchers have also been trying to use neural networks to simulate the human brain to generate new data.
Unfortunately, so far, this data is often very problematic and completely unusable.
And a big resistance is that the generated images are often indescribable. These useless grammatical images lack the means to exploit them, and it is difficult to evaluate them.
A few days ago, he got new inspiration when the model performance on IMAGENET exceeded human standards.
If deep networks can already reach human levels on some image tasks, can we avoid the problem of difficult evaluation of generative network results and complete this process through the mutual confrontation of the two networks?
The more Ian thought about it, the more reasonable it became. He didn't even bother to sleep last night. Instead, he spent the whole night implementing a prototype version of generative-confrontation, and tried training it, and found that the results seemed to be basically as expected. consistent.
Unable to contain his excitement, he wanted to discuss the principles and methods with his instructor as soon as possible.
Ian wandered back and forth at the door of Bengio's office for more than twenty minutes before Bengio strode over from a distance.
If you ignore the white color in his hair and beard, the 49-year-old Bengio actually looks quite young. Judging from his appearance, he could be a man in his mid-thirties at most.
In particular, the two thick eyebrows that look a bit like Mr. Bean make him look more loyal and approachable.
Ian quickly took a few big steps forward, and at the same time unfolded the notes and experimental records in his hands. He pointed to the core formulas in the above lines, and when he was about to speak, he was ruthlessly interrupted by Bengio.
"I know you are in a hurry, but don't be in a hurry yet." Bengio looked at Ian's irritable look and joked, "What we pay attention to in academics is to have a solid foundation and work steadily. What's the use of being anxious? There's no point in being anxious. "
"What can't we talk about after we get in?" Bengio said, taking out his key and opening the office door. "Come in, it's just past seven o'clock and you're squatting in front of my office door. You must have made some important discovery, right?"
"It's urgent, I'm really anxious." Ian has been deeply involved in the field of generation for some time. When he thought of this idea yesterday, he had already realized that this would be a major change in the entire model.
Ian is very excited to have spent ten years sharpening his sword since he was an undergraduate. He has a good academic sense and can sense the extraordinaryness of this idea. Is your day coming soon?
At this moment, he could feel that this was an excellent opportunity to rely on his own strength to get rid of the halo and shadow experience of Ng Enda's disciples and Bengio's disciples.
Let the world truly understand himself and re-see himself through the technologies and algorithms he proposed.
"Are the two networks fighting each other?" Bengio pondered for a while after listening. Unlike several of Ian's friends who were eager to raise objections and doubts, Bengio has always been interested in imaginative and creative ideas. All very encouraging attitudes.
“Train two models at the same time instead of one. The generator and the discriminator can spirally improve each other with left foot and right foot. You want to use this model to make the distribution of the generator as close as possible to the provided real world. data."
"In this way, we also get rid of the ordinary thinking. We train the discriminator not so that it can judge 100% whether the image is generated. Instead, our purpose is to make the discriminator no longer unclear about which ones are true and which ones. It’s fake,” Ian added.
"Confrontation is just a form, and its purpose is to train a generative network. I am confident that this method can get rid of the dependence on data labels, and a large amount of other data can be utilized without additional processing. stand up."
Until now, Ian's excitement had not subsided.
"Not only that, generation is only the most preliminary application." Bengio thought for a while and said, "It can achieve special purposes by filtering data. It only requires a small amount of code adjustments, and I believe it can be very good. It is a very dangerous technique to process patterns with simple layouts, especially in the direction of human faces."
"What are you going to name it?" Bengio said as he turned on the computer. Naming technology is a difficult problem and an art.
If you choose a bad name, your work will be in vain.
A catchy name, an interesting and concise abbreviation, are all likely to increase the visibility, popularity and application scope of the work.
"Generate...and...confrontation. How about calling it a generative adversarial network?" Ian added the two most critical elements of this idea to the name before he thought about it carefully.
"Generative Adversarial, let me take a look." The first step in choosing a name is naturally to search to see if anyone has already occupied the slot.
After Bengio's search, Meng Fanqi, who was far away, suddenly sneezed and thought to himself, how could he still catch a cold in this summer?
"This name seems to have been taken by someone, just two days ago, which is very unfortunate." Bengio did not realize the seriousness of the problem at this time, "It seems that we have to change to another name."
As he spoke, Bengio clicked on the preprint.
"What kind of network is DreamNet? I haven't seen it before." Bengio vaguely felt that someone had mentioned this name in the past two days, but he was busy with other research matters and didn't pay much attention to it.
But just after reading the summary, Bengio's eyes sharpened and his expression became very serious.
Ian was still thinking hard about what kind of name would be better.
And Bengio ruthlessly broke through his defense with one sentence, "Ian, it seems that it's not just a name we can't use."
"What? Our ideas crashed?" Ian was shocked. Academic ideas crashed, which may be one of the most unlucky things for scientific researchers.
