Deep Space Fleet

"Please ask the two players for 3 minutes."

After a short while, the host passed the microphone to Cai Kun.

"Please ask Cai to answer below."

He blushed and stumbled for a long time: "Sorry, I haven't delved into this field ... A recurrent neural network is an artificial neural network with a tree-like hierarchical structure and network nodes recursively input information according to their connection order. One of the deep learning algorithms. "

I barely memorized some definitions, but I was a bit confused. Classmate Cai said, "I'm sorry, I don't have any particularly good ideas for the time being."

There was a boo from the bottom of the floor. Although student Cai was blushing, he still stood on the stage stubbornly.

He looked up at Zhang Yuan.

It depends on what you can answer. If you ca n’t answer anything, why is your interview score higher than me?

"Someone asked Zhang to answer below."

"I have a question." Zhang Yuan raised his hand, "I have heard of the Hopfield Network, but what is network stability?"

The questioner smiled: "If some weights of the network can converge to the equilibrium point, that is, the weights are said to converge, then the output can fit the expected output, that is, the system is stable, so the convergence is for the variables In terms of stability, it's about the system. "

"The stability of the system cannot be guaranteed, the control system is unstable, and the network convergence has lost its foundation."

远 Zhang Yuan pondered for a moment, and said, "I haven't researched your research field in depth, just read some papers. But I have some simple ideas that you can listen to."

"... First is the monotonicity problem. In the Hopfield network model of discrete-time continuous state, is the activation function of a neuron a monotonic function, or whether it is locally monotonic."

第二 "The second is to study the conditions under which the energy function becomes a convex function, and consider the operation of the Hopfield network as a solution to a constrained convex optimization problem, and then try to demonstrate whether there is a sufficient condition for a globally unique minimum point ..."

"Convex optimization, you are right! Convex optimization is the direction I am doing, and I have another question ..."

一次 This time, it comes to the mind of this questioner, he raised several questions about convex optimization, and exchanged mathematics on the spot.

Finally, under the direction of the host, Zhang Yuan had to say: "For the knowledge about convex optimization, if you need it, you can come to me to discuss it behind the scenes, and it will not be explained in detail here ..."

A lot of things, he just talked casually based on his usual accumulation. If he could really write a dissertation at random, he wouldn't have to go on stage, and he would be a "thesis god".

There was applause underneath me.

For such a short period of time, the verdict was made.

Cai Kun could not wait to find a slit in the ground to get in.

He could only blush and secretly cheer himself up, perhaps because of the difference caused by too few samples, and he just encountered the problem that he would not, the other party would.

The host asked again, "No. 666, please speak your question."

The questioner is a girl. Her question is relatively pragmatic and not so partial: "I want to ask one about the server, the cache."

"The cache is divided into two phases: the first is the data placement phase, during idle periods when the data demand is small, using the available communication resources, the data is placed in the cache device of each user. The second is the data distribution phase, Assume that during the peak demand for data, each user randomly requests a complete file from the server. The server takes these needs into consideration and distributes the complete data to meet the needs of all users. "

"My question is, how to design the caching scheme most scientifically?"

The questioner even sent a PPT to the screen.

问题 This question is very professional, but everyone can understand what is being asked.

The host said, "Okay, everybody, please think for 3 minutes."

The microphone was delivered to Zhang Yuan.

He said with a smile: "I think ... let's give the opportunity to Cai Kun first. I'm afraid he didn't say anything after I said it."

The people underneath laughed and whistle broke out.

Some people believe it, others don't.

Cai Kun thought for a moment, this time, he still had a little idea, but Zhang Yuan on the opposite side seemed to be unable to say anything and needed more thinking time.

"... I have the idea that each user caches the data of the M / N ratio of each file separately, and during the data distribution stage, the server releases the data of each (1-M / N) part that each user is missing, one by one, At this time, the value of the transmission data is R = K (1-M / N) ... "

"I don't know if you have heard of a scheme called" coding cache ", the specific algorithm is like this ..."

"You make sense."

Although this is said in the mouth, the questioner is slightly disappointed. The coding cache is arguably the most trivial idea. There are already mature algorithms already, and it has no research significance. If this bad idea is used to start a project, it is estimated that the instructor will slap it back.

Next Zhang Yuan's turn, he cleared his throat.

"Cai Kun's idea is very good. During the distribution phase, the existing cached information is used to design a certain encoding combination of the required broadcast content so that multiple users can decode it from a single message at the same time. Needed information to get global cache gain. "

"But its limitation lies in ~ www.readwn.com ~ It is necessary to divide each file into a parameter that grows exponentially with the number of users K. It is well known that the index-level division is difficult to achieve algorithmically ..."

"May I borrow your digital pen and big screen?"

The host froze, "Yes."

Zhang Yuan drew several pictures on the screen.

"... In essence, it is a problem of combinatorial graph theory. What we want is a cache scheme when F is a polynomial level of K and R is a constant level. Or prove the existence of such a scheme."

显然 "Obviously, whether the number of file divisions can be converted into a polynomial level, that is, a level that can be used in industry, depends entirely on the mathematical ability to construct a hypergraph problem that meets certain characteristics, these characteristics are like this ..."

"I think I can write a good dissertation after turning it into a mathematical problem. As for what the answer is, I can't answer it now, it's just an idea. How to solve this graph theory problem? Think for yourself. "

There was a whisper of whispers.

There are still many professionals who eat melon here. Is it nonsense, naturally it has its own discrimination.

Crackling!

After a few minutes, the questioner took the lead and applauded.

There was thunderous applause.

Cai Kun has no face to continue, he flushed, and nodded hurriedly to give in.

终于 He finally knew that the same 100-point test paper, he took 100 points because of strength, while others also took 100 points, because there were only 100 points on the paper.

"Lao Cai, you have encountered a real metamorphosis!"

"Yes……"

"It doesn't seem to be wrong."