Many schools will have an exhortative motto, similar to "Today you are proud of the school, tomorrow the school will be proud of you."
To most people, this is just chicken soup. But there are a small number of people who can really make the school famous on their own.
The same is true for papers. Some papers are proud to be published in top journals and conferences, while some papers are the reason why journals and conferences become top.
People who know a little bit about scientific research will basically think that good papers should be published in SCI, high-scoring, and first-tier SCI journals. There is indeed nothing wrong with this view most of the time.
However, for computers, especially in the AI field, conferences are actually much more important than journals.
The conference is a form of receiving manuscripts at regular intervals every year, reviewing manuscripts at regular intervals, and then bringing all the authors of successful manuscripts together to communicate and present. His characteristic is that he reviews manuscripts quickly and his time is fixed and guaranteed. And participants can have direct face-to-face discussions with the authors.
Journals, on the other hand, can submit manuscripts at any time throughout the year, the review time is not fixed, and the traditional format of meeting with authors is not organized.
For many disciplines, journals are more authoritative, more formal, and more rigorously reviewed.
But for AI, journals are too slow. The speed of AI is so fast that many people can’t wait.
For example, when AlexNet won the championship last year, you did a lot of research yourself at the beginning of 2013, studied it for four or five months, and then submitted the manuscript. If you apply for a conference, the review results will have been announced in August and September.
In November or February, you can go to the venue comfortably, chat and laugh with others, and then graduate successfully.
But if you submit to a journal, you may still be reviewing the manuscript after half a year. At this time, as soon as the new results for 2013 come out, the reviewer takes a look and finds that the reviewer seems to be several points higher than you.
It is normal for an article to be rejected because it might have been successful.
At this time, it can be said that Tiantian Yingying and Didi are no longer working. If you are lucky, you may switch to a worse conference and journal, and it may still go smoothly.
If you are unlucky and your article is rejected 2-3 times, it is very likely that it will be completely outdated from now on. Half a year of hard work will be in vain. For many students, it may be the end of a one-year delay in graduation.
Due to this unique and fast disciplinary pace, the AI direction emphasizes conferences but not journals. Often, only when success is achieved at conferences will the article continue to be expanded and submitted to journals for review.
If scholars in the field want to understand the latest trends, they will basically not check journals. Instead, they will pay attention to the top academic conferences and attend the conferences to communicate directly with the authors.
Journal? When it was published, the technology was already a year and a half ago. Times have changed, my lord.
But at this moment, even the meeting was still too slow for Meng Fanqi. The results of IMAGENET have been initially announced, and he believes that interested people who are paying attention to this matter have set their sights on DreamNet and the team Dream.
Now that the New York Times has paid attention to this matter, it seems that it is time to launch a generative adversarial network to add fire to DreamNet.
Some articles rely on the reputation of high-level conferences and journals to support themselves, while others do not.
The deadline for receiving CVPR submissions at the AI Summit is November 1, 2013. After Meng Fanqi submits his submission, he can choose to make his submission anonymously public.
It is not against the rules to do so.
It's just that sometimes, this matter is just covering one's ears and stealing the bell, taking off one's pants and farting.
For example, in the current situation, Meng Fanqi is preparing to release his preprint, "Generative Adversarial Network Based on DreamNet".
Even if he chose to remain anonymous, who wouldn’t know who the author of this article would be?
No information about the current details of DreamNet has been released. Who else but him could publish this paper based on DreamNet at this time?
In the same way, later generations of AI technology are gradually developing towards large models, often requiring hundreds or thousands of GPUs. Many times, you can guess which company and which research group it is by just looking at how many cards are used and whose unique big data is used.
Others who want to do it don’t have the resources or the ability.
Since it was unnecessary, Meng Fanqi would naturally not do such a boring thing.
"At this moment, what I'm most afraid of is that others won't recognize me, so why bother being anonymous?"
With this in mind, Meng Fanqi published this paper on the arxiv website. As for whether people at the CVPR meeting will feel that this approach violates the principle of double-blind review, that is not within the scope of his concern.
arxiv is an open academic preprint storage website and a modern way to share research results.
At first, it was because it took too long to review manuscripts in some basic subjects, especially mathematics and physics. Some papers may not be read by anyone for several months, and no one can understand them at all.
In this case, many people will consider putting a draft version, or simply the final version, directly on the public platform arxiv.
This can promote communication and the development of the discipline, and is also a kind of evidence to prove when your research results were obtained.
Later, computer, statistics, biology, economics and other disciplines were gradually added, and arxiv became more and more all-encompassing.
Unlike serious conferences and journals, arxiv is just an open platform. It’s the same as pixiv, an illustration sharing website. arxiv and pixiv basically do not conduct very strict review of uploaded content, so the content level above varies.
arxiv is not a formal conference or journal, and its content is not really published. Content posted above has not been peer-reviewed and is at risk of non-recognition.
Minke could also publish its perpetual motion masterpiece on it and pretend to be inscrutable.
If nothing else, it does wonders to trick the unsuspecting layman. The paper introduction, various field labels, and citation methods on the page are all very presentable, and they are all in English. Non-scientific researchers can easily be fooled at first glance.
But Meng Fanqi was not worried about this matter at all. The logic of modern AI is so simple and crude.
The principle and code are here, and with the same random seed, anyone can reproduce my results.
Meng Fanqi didn't care at all whether others admitted it or not, questioned it or not.
In fact, he was vaguely expecting it, because this kind of conflict would bring considerable attention.
Not long after, Meng Fanqi submitted the latex source file of the generative adversarial network and compiled a pdf file on the website. After only two days, the website will update these new articles.
At the same time, the original author of the paper, Ian Goodfellow, has just recently thought of the prototype of this idea and is still in the thinking stage. There is still a long way to go before the truly formed generative-confrontation framework.
Ian is on his way to the office of his doctoral supervisor, Bengio, one of the three giants of AI.
“This is a great idea and I need some help,” thought Ian. “I’m going to spend the last year of my PhD doing the best work I’ve ever done.”
He would never have imagined that a more complete article, with more detailed experiments, and more comprehensive discussions and reflections on related fields, would have been published.
