I control the AI ​​empire thanks to the geniuses in my group.

The filling materials, experimental data, methodological details, model architecture, etc., were all things he created himself; no one knew them better than him.

Su Zhou turned on his computer, created a new LaTeX document, and configured the format according to the IEEE two-column layout template.

Then, he made an outline on the draft paper.

After outlining the points, Su Zhou took a deep breath and began to write.

To be honest, Su Zhou got stuck when writing the first paragraph.

The thoughts in my mind are like a tangled mess of headphone wires; I know there's something inside, but I just can't seem to pull it out.

But after he typed the first word, the rest of the content flowed out like water.

He knows his work all too well.

He knows exactly how every line of code is written, how every model is trained, and how every set of experimental data is generated. These things are ingrained in his mind.

The only obstacle was the expression in academic English, but this problem was smaller than he had imagined.

After spending some time with Newton, his English reading and writing skills had already been honed through repeated practice.

The language of a paper doesn't need to be fancy, it just needs to be precise.

This is precisely Su Zhou's strength.

For the next few days, Su Zhou lived a double life that could be described as self-torture.

During the day, I'm at school preparing for the monthly exams.

When Su Zhou got home that evening, she locked herself in her room and continued writing her thesis.

The method description section is the most time-consuming. It requires precisely describing the structure of each layer of the network, the forward propagation process, the design of the loss function, and the gradient update strategy using mathematical language.

He scrutinized each formula repeatedly to ensure there were no logical leaps.

He kept Turing's words in mind, and the reviewers would look for flaws from every angle.

Then he would plug the loopholes in advance.

The experimental section was the part that Su Zhou wrote most smoothly.

After all, the data is readily available.

His detection accuracy and inference speed on public datasets such as PASCAL VOC far exceeded those of mainstream methods at the time.

Once the comparative experiment table is presented, the numbers themselves serve as the best proof.

The ablation experiment was also conducted according to von Neumann's suggestion, removing the multi-scale feature fusion module, attention mechanism, and adaptive anchor box strategy one by one, and observing the decrease.

The results show that each module makes a significant contribution.

By the sixth day, the first draft was finally completed!

Su Zhou read it from beginning to end, and then began to revise it.

Shannon was right; good papers are the result of revision.

He revised the wording three times, the formula layout twice, the chart color scheme once, and finally checked the reference formatting to ensure consistency.

The text consists of eight pages of densely packed double-column layout, four architecture diagrams, six experimental result tables, and three visualization comparison charts.

carry out.

Su Zhou stared at the PDF on the screen and remained silent for a long time.

This was the first academic paper he ever wrote.

Thirteen years old.

the first time.

He had no professors as his mentors, no large research group, and no high-end laboratory.

His laboratory was the computer he assembled in his bedroom, and his mentors were a group of geniuses from different eras in Group 42.

Su Zhou saved the paper as the final version and opened the online submission system for CVPR 2008.

CMT's interface loads in the browser; the blue-gray page design is so simple it's almost rudimentary.

Su Zhou registered an account, filled in her personal information and thesis information, and then uploaded the PDF.

The progress bar moves slowly.

12%……36%……58%……87%……100%。

Submission successful.

Your paper ID is: CVPR-2008-3847

You will receive a confirmation email shortly.

Su Zhou stared at the green success message on the screen and let out a long breath.

It's been thrown.

Regardless of the outcome, this step has been taken.

A few days later, at the Computer Vision and Machine Learning Laboratory, Technical University of Munich, Germany.

Professor Klaus Schmidt sat at his large desk, with three printed papers spread out in front of him and a cup of cold black coffee to his right.

His gray hair was neatly combed, gold-rimmed glasses sat on his nose, and his gray-blue eyes revealed the sharpness cultivated from years of academic work.

Professor Schmidt is a senior reviewer at CVPR, receiving ten to fifteen review invitations every year. In the field of computer vision, especially in object detection and image segmentation, his name is almost universally known in European academic circles.

There were three manuscripts awaiting review on his desk, all assigned by the editorial department of CVPR 2008.

The first one is from MIT, about image segmentation.

The second article is from CMU and is about feature extraction.

The third article is from...

Professor Schmidt squinted and glanced at the author information.

Zhou Su. Chongming High School, China.

His brows furrowed almost instinctively.

"High school?" he repeated softly, thinking he had misread it.

You read that right.

It was indeed a high school.

He then looked at the title: "A Deep Neural Network Framework for Real-Time Object Detection and Scene Understanding".

Neural networks, real-time object detection, and scene understanding.

Professor Schmidt leaned back in his chair, took off his glasses, and wiped them.

He understands this area.

The prevailing view in academia is that deep neural networks perform far worse than traditional feature engineering methods in computer vision tasks.

While convolutional networks perform well in handwritten digit recognition, scaling them to more complex visual tasks presents significant bottlenecks in terms of computational resources and data volume.

An author from a high school in China, working on deep learning all by himself?

Professor Schmidt's lips twitched slightly, whether in mockery or confusion was hard to tell.

"Marcus." He looked up and called out to the student who was passing by the door.

"professor?"

Professor Schmidt picked up the paper from the table and handed it to him.

"Please give this a preliminary review."

Marcus took it, glanced at the title and author information, and his expression was exactly the same as Professor Schmidt's.

"Mingde High School? This is... a middle school?"

"It seems so." Professor Schmidt shrugged. "It could be a system error, or it could really be a submission from a high school teacher. In any case, let's see if it has any basic academic value. Don't spend too much time on it, just give it a try."

He picked up the glass of cold coffee, took a sip, frowned, and put it down.

"Chinese people don't have much experience in deep learning yet. Most of the papers submitted to top conferences are just copying others and lack originality. I've never heard of the author of this paper before, and his workplace is a middle school. He's probably just filling in the numbers."

"Understood." Marcus nodded and put the paper into his folder.

Professor Schmidt glanced at the title of the paper again.

Real-time target detection and scene understanding based on deep neural networks.

Ha, what a boastful tone.

Professor Schmidt smiled silently, picked up the now-cold black coffee, and forced down another sip.

The computer vision community is becoming increasingly superficial; anyone can submit papers to top conferences like CVPR.

It's 2008 now. What's the consensus in academia?

The top scholars are dedicated to manually designing feature extractors using more elegant mathematical models and deriving the relationships between pixels through rigorous probabilistic graphical models. This is the right way to truly test a researcher's mathematical and algorithmic skills.

As for deep neural networks?

That was just a dead end that was hyped up in the 1980s and 90s, and was later proven by theory to have fatal flaws.

"Real-time object detection..." Professor Schmidt shook his head.

I went back to reading the MIT paper and didn't think about it anymore.

In his many years of experience reviewing manuscripts, he has seen far too many submissions of dubious origin.

Most of them don't even pass the initial review.

P.S. It's the beginning of the month, so the author is asking for more reading, monthly votes, and recommendations!