178: Chapter 178 It was him?!

Charles soon reached the third part of the paper.

[At this point, I hope readers will remember our domain name, the Taiwan novel-chasing artifact kan.com, to read at any time.]

In this section, the paper demonstrated a distributed deep learning training process simulating 100 computing nodes and 10,000 iteration cycles through a code-constructed virtual environment on a local computer.

In an extreme asynchronous environment, varying degrees of network latency were added to each computing node.

Yet, it ultimately achieved a very good result.

Charles stared at the final convergence graph for a long while.

On the graph, the red line representing the error rate did not exhibit various errors or crash directly after several thousand iterations, unlike traditional asynchronous models.

Instead, that red line showed an extremely stable downward trend. The final error rate was controlled within a reasonable range, and the entire model did not crash.

Very stable!

Charles took a deep breath, excitement showing in his eyes.

"My God..."

Charles leaned back in his chair, muttering to himself.

As a senior reviewer and editor for the NIPS conference, he also had his own computer laboratory at the University of California, San Diego, and possessed high professional standards.

Because of his professionalism, he saw at a glance the immense value hidden behind this brand-new architecture.

This was not just a theoretical derivation paper on paper. The author also provided irrefutable experimental data.

This kind of data was impossible to fake because it was too easy to reproduce; the possibility of data falsification was extremely low!

If this asynchronous framework based on dynamic timestamp weight penalties could achieve industrial-grade deployment...

Thinking of this, Charles's breathing became a bit heavy.

If it could truly be applied, the communication bottleneck currently plaguing the entire AI academic community would be broken.

The training scale of neural networks would break free from the physical limitations of single-machine computing power or synchronization locks, ushering in an explosive growth.

This paper was of great significance! Its academic value even exceeded all accepted papers for this NIPS conference.

At this thought, Charles excitedly grabbed the landline phone on his desk and quickly dialed the number of the Executive Chair of the Main Program Committee of this NIPS conference, a senior professor in the Department of Computer Science at Stanford University.

The call was quickly answered.

"Hello! Mr. Chair! This is Charles."

Charles shouted, his voice tinged with excitement.

The Executive Chair on the other end was clearly handling other matters and was somewhat stunned by Charles's rare loss of composure.

"Charles? What's wrong? What happened to make you lose your composure like this? Has the Google Brain team submitted some new results regarding large-scale unsupervised learning again?"

The Chair asked curiously.

"It's not Google! Nor is it Microsoft or the team from the University of Toronto! It's a completely unknown individual contributor named Yang Zhao!"

Charles held the receiver and said loudly, "He solved the gradient staleness problem of the asynchronous SGD framework that has been bothering us! He introduced an extremely advanced model based on latency time differences to assign dynamic weight penalties."

"He achieved stable model convergence in a completely asynchronous environment! Moreover, the experimental process and data are provided in the appendix!"

"Are you sure the derivation process and data aren't faked?"

Hearing that the gradient staleness problem had been solved, the Executive Chair's tone immediately became extremely serious. This concerned the future development direction of the entire distributed computing field.

"I spent an hour roughly going through his core algorithm and mathematical derivation formulas line by line. The logic is seamless, without any obvious loopholes. As for the data, I personally believe the probability of falsification is extremely low."

"Because the difficulty of reproduction is very low."

Charles answered quickly, "I've already sent the full paper to your work email. Take a look for yourself, quickly!"

"Alright, I'll download and check it immediately."

The Executive Chair said directly.

"Immediately set this paper's priority to the highest in the backend and add it to the Grade A priority review channel. I will personally contact three of the top distributed computing experts from MIT and Berkeley for an emergency blind review!!"

"Understood, I'll do it right away."

Hanging up the phone, Charles sat back down in front of his computer.

In the NIPS internal review system, he changed the status of this paper to the highest level of expedited review and distributed it to the several external independent reviewers designated by the Chair.

Having done all this, Charles looked at the PDF file still open on the screen, his gaze falling on the author's signature and affiliation below the title.

"Yang Zhao... Beijing Forestry University? Peking University? Just where did this monster come from?"

Charles muttered to himself, his eyes full of curiosity. He had been in the circle of top scholars in the field of computer science for over a decade and knew many Chinese scholars from various prestigious universities, but he had no impression of this name or this institution.

Hesitating for a moment, Charles opened his browser and entered the Google Scholar search interface.

He typed "Yang Zhao" into the search box and added "Beijing Forestry University" as an institutional filter for an accurate search.

The page jumped, and the search results popped up.

Ranked first was a paper on image denoising published in IEEE TIP (IEEE Transactions on Image Processing).

"A top journal in the field of image processing? It seems he really is a scholar in the direction of computer science." Charles nodded slightly; this was a reasonable background and resume.

But when his gaze moved down and saw the second paper in the search results, which had just been indexed by the system...

Charles's hand holding the mouse froze.

Paper Title: On the Boundedness of Prime Gaps.

Journal: Annals of Mathematics.

Author: Yang Zhao.

Although Charles was in computer deep learning, he was also a scientific researcher. He knew very well the status of the Annals of Mathematics in the scientific community, and even more so what the four words "prime gaps" represented.

Just two days ago, Princeton University's official website and various science media outlets were still reporting overwhelmingly about a young Chinese scholar who had proven the twin prime conjecture.

He had heard a bit about this news.

Extreme shock appeared in Charles's eyes, and his pupils contracted slightly.

"My God... the person who solved this asynchronous communication algorithm is actually that mathematician who just proved the twin prime conjecture not long ago?!"

Charles stared at the screen, muttering to himself in disbelief.

"He also published this TIP paper. A month ago, he published in a top image processing journal; a few days ago, he published in the top journal of the mathematics world to solve a century-old problem; and now he has submitted a computer paper to NIPS that is enough to change the underlying architecture?"

This extreme interdisciplinary ability and terrifying output speed completely overturned Charles's understanding of the limits of human Intelligence.