Spark

The things on the sea quickly reached Chen Nian's ears, but he was no stranger to such a showy operation.

Isn't it the basic operation of all countries to use warships that are supposed to be decommissioned for fishery administration and maritime police?

It's just international practice, nothing to be proud of!

What's more, the 051 used for this transformation is already old and not good enough. According to his memory, he will retire in 2006. Before retiring, it is nothing to use the residual heat to temporarily save the field.

However, for him, after this incident, the construction of the island was put on the agenda, the obstacles were basically removed, and the remaining work was to build our own Sky Whale in the shortest possible time.

Chen Nian did not go to the scene to direct, because compared with all previous projects, the overall technical difficulty of the Sky Whale was not as high as imagined.

Its difficulty is mainly reflected in the difficulty of engineering.

In other words, the difficulty lies in how to dismantle and distribute a huge project in an orderly manner, how to mobilize the productivity of various departments, and how to carry out streamlined and standardized production.

In these respects, Chen Nian actually couldn't help much.

All he can do is to use his own strength to help the project team break through quickly when they encounter technical obstacles.

So, it doesn't make much difference whether you are on site or not.

Therefore, during this period of time, Chen Nian stayed at Xigong University to continue his studies.

He has finished all the courses of advanced engineering materials, but unlike those written in those fantasy novels, after learning a complete cheat book, he did not immediately obtain a powerful kung fu.

Because the accumulation of this thing is done step by step, there is no chance of reaching the sky in one step.

But looking back, this entire period of study provided him with more than 15 source points, and overall, it was extremely cost-effective.

In the next step, Chen Nian intends to continue to study material surface/interface science to further complete and enrich his knowledge system of materials science.

The course has been booked, and the instructor is still Mo Zhongmin. With his level, he can teach the current Chen Nian very well.

Apart from attending classes, what Chen Nian does most now is to discuss the issue of the drone system with Wang Jiancheng.

This was originally Wang Jiancheng's project, but after seeing the potential of the project, Chen Nian couldn't help but get involved.

It's not that I want to pick the fruit of Wang Jiancheng, but it's just out of a scientific researcher's curiosity and desire to conquer.

He wondered if he was capable of producing breakthrough results without the help of the system.

This kind of thinking is the same as the elites in the high-ranking workplace who continue to apply for jobs naked.

It's all to test your true strength, to avoid boiling the frog in warm water, and finally dying in the pot.

So far, the progress has also surprised him.

Although Wang Jiancheng is indeed an out-and-out genius in this regard, his guidance and assistance also played an irreplaceable role.

After fully understanding the principles of the convolutional neural network algorithm, it took the two of them less than three weeks to create the first version of the weak artificial intelligence training model, ready to start preliminary testing.

As an important project partner, Chen Nian will naturally not miss this significant test.

After finishing a small class for a day, Chen Nian and Wang Jiancheng made an appointment in the laboratory of the school's comprehensive computer education center, where there is a minicomputer capable of supporting the computing power required by the convolutional neural network model, which can be used for this training experiment.

——

Of course, being approved was actually the result of Chen Nian's secret arrangement through Chen Guo.

Otherwise, although Chen Nian himself could come in and out at will, but as Wang Jiancheng is an ordinary student, no matter how awesome he is, it is impossible for him to open this thing.

Sometimes Chen Nian couldn't help feeling a little emotional.

Scientific research is like cultivating immortals, wealth, law, and land are indispensable.

He walked all the way to the laboratory. After entering the door, Wang Jiancheng had been waiting there early. When he saw him walk in, he immediately stood up and greeted him warmly:

Senior! Over here!

Chen Nian went over and saw that not only Wang Jiancheng was present, but also Chu Jianshu, a professor of flight control at Western Polytechnic University who he had asked for help before.

At that time, Wang Jiancheng was going to start the project, but the resources and information preparations were not sufficient, so Chen Nian contacted Chu Jianshu through Chen Guo's relationship and asked him to be an interface person.

It is also through his relationship that this laboratory equipped with a minicomputer will be opened.

The other party was not a member of the Xinghuo team, and he didn't know Chen Nian's identity. In his eyes, Chen Nian, like Wang Jiancheng, was an ordinary and relatively talented student.

He hurried up to say hello.

Professor Chu, are you here?

Chu Jianshu nodded with a smile, and replied:

Your project is very interesting. I followed it before, and now that it has produced results, I must come and see it. Doesn't affect you?

It doesn't affect it, it doesn't affect it. If you are willing to guide, we are too happy.

Chen Nian replied with a smile.

On the other side, Wang Jiancheng has already set the basic parameters of the model and connected to the flight simulation program.

In the subsequent training, the model will automatically start learning according to the random instructions input by the flight program, and use high-frequency trial and error to gradually find the most profitable strategy.

So how did you set your criteria in the end? Is it still the same as I said before, with the following distance as the standard?

Chu Jianshu asked.

No, it has been changed. This criterion is too rigid. If only distance is used as the criterion, although the success rate is high, at the same time, once it fails, the consequences will be very serious.

We can't add the estimated loss as a criterion. In that case, the whole model will become too large.

Even if our small computer can run now, it won't be able to run with a real drone in the future.

So, the final strategy is to use the success rate of returning to the team as the main weight criterion, and the distance is only used as an auxiliary.

Hearing Chen Nian's answer, Chu Jianshu's eyes lit up.

This is indeed a very clever strategy.

As an important part of the formation, there is no doubt that UAVs must form a flight formation with manned aircraft, but due to differences in performance, the formation of the formation cannot remain stable forever.

——

Let alone manned and unmanned formations, even purely manned and machine formations cannot maintain their formation in battle.

But on the other hand, if you want to form combat effectiveness, formation and grouping are necessary.

This requires the pilot to return quickly after completing the tactical maneuver.

Using this as a standard to restrict the actions of drones and use it as a criterion for machine learning training, on the one hand, it gives the model more freedom, and on the other hand, it can also meet the needs of actual use. It can be said that it is a compromise. The best choice ever.

These two boys are indeed geniuses, they can master everything in one way, and for this model, they definitely have more than just a single knowledge of artificial intelligence.

Object-oriented development is really useful development.

However, how could an ordinary student have such insight?

——

He still doesn't know that at this time, Chen Nian's understanding of air force tactics has actually been experienced in practice

Clever, this should be our best solution. Shall we start now?

Start now!

Wang Jiancheng nodded and tapped the Enter key with his finger.

With the fan of the minicomputer spinning, the program started running immediately.

A piece of data is continuously input into the model. Since there is no visual interface, all the learning process is presented in the form of data.

Chu Jianshu didn't quite understand it, but Wang Jiancheng, staring at the frantically flickering lines of code, kept explaining:

A thousand times of training have been completed, and the effect is very poor. The success rate is 0%, but it is normal. I estimate that it will take at least a million times of training to achieve the initial effect.

Five thousand times, this is considered a success. One drone returned to the team, but this is luck, and the strategy has not yet been formed.

Time passed by, and as the program continued to run, Wang Jiancheng's face became more and more serious.

There is no other reason, the success rate is too low!

The training effect of this model is far weaker than he expected. If this efficiency is followed, training at the level of 100 million may not be enough.

Are you still playing a fart?

There is an iron law in machine learning:

When the training intensity of AI is greater than that of artificial training, the model is invalid.

Obviously, the model in his hand is like this.

Not very good. This algorithm does not seem to meet the usable standard.

The training efficiency is too low, and the speed of strategy generation is also slow.

If we go on like this, we might not be as cost-effective as manually traversing the custom strategy.

Hearing his words, Chen Nian also frowned.

He asked:

Can you see why?

I can't see it. I feel that there should be no major problems with our algorithm, but the cold start consumes too many resources.

The essence of machine learning is also traversal, but it cannot make any reference to perceptual experience.

That is to say, artificial priors and machine priors are inconsistent

Looking at Wang Jiancheng's slightly disappointed eyes, Chen Nian continued to ask:

What about the confusion matrix we talked about last time? Can it be solved?

It can't be solved. Now our problem is that there are relatively few priors and experience. If we want to get high-level solutions efficiently, we must require a very high level of intelligence in the machine.

And at present, the generalization of this task, that is, the difficulty of task generalization itself is very high, which further increases the requirements for machines.

No, it's not a problem to run like this.

As he said that, Wang Jiancheng directly hit the stop button, and the model stopped running. In the end, in more than 60,000 training sessions, only two valid results were achieved.

This amount of data is simply not enough to support the model to continue training.

Wang Jiancheng looked at the results on the screen in disappointment. At this moment, he felt an indescribable sense of frustration.

In fact, at the very beginning, he was extremely confident in the model.

After receiving Chen Nian's instructions and gradually understanding the basic principles of the convolutional neural network algorithm, he almost felt that the whole world was in his hands.

The potential of this set of algorithms is really too great. As long as we continue to develop it, it is not a luxury to change the world completely one day.

Moreover, on this track, I am already many steps faster than other researchers. As long as I can produce usable results and cases, I am likely to become one of the pioneers of the new generation of artificial intelligence together with my seniors.

What an honor this is, needless to say.

With such self-confidence, he can be said to be advancing all the way, so fast that even Chen Nian was amazed.

He originally thought that things would go on smoothly like this, but he didn't expect that the initial training results of the model gave him a heavy blow.

It's not just that the result is not as expected, it's not usable!

If it fails to meet expectations, it can be continuously adjusted and optimized to get closer to the goal, but if it is not available, it is really useless to try.

It might be the only way to start over.

But can I still have that kind of spirit?

Chu Jianshu silently listened to the conversation between the two. In fact, this result was within his expectations.

You know, this is artificial intelligence!

Even in science fiction, the prospect of artificial intelligence is decades away.

But in this era, in order to realize basic artificial intelligence under the current computer computing power, the performance limit of the terminal must also be considered. How easy is it?

Being able to do this is already a success in a certain sense.

The next road still needs to be walked slowly step by step.

——

Of course, when it comes to solutions, he is not without them.

Before, because of his interest in the project, he had also learned some knowledge about machine learning and convolutional neural networks.

According to the discussion between Chen Nian and Wang Jiancheng just now, the biggest shortcoming of the whole model is the lack of prior experience. If a large enough sample database can be embedded in the model, and the machine can perform deduction and learning according to the samples in the database, then the machine training The difficulty should decrease geometrically.

Such a strategy is not easy for others, but it is not difficult for him.

If you want data, it's just a matter of a few phone calls.

Based on your own scientific research purposes, wouldn't it be enough for the military to send some of the pilot's training records? If not, let them fly a few times according to the target logic.

But the problem is that such a solution to the problem is tantamount to counterproductive for these two children.

This time, I can help them solve their problems through privileges, so what about next time?

If this is an urgent scientific research task, no matter what method is used, Chu Jianshu will not feel that it is too much.

But right now, it's more of a quest.

Talent needs to be cultivated. In the process of their leveling, you can give them some help, and you can also show them the way, but you must not help him defeat the boss

Thinking of this, Chu Jianshu asked calmly:

Then what are you going to do now?

The two were silent for a moment, then Wang Jiancheng said:

There are two ways to go.

Or, I will find a way to contact the aircraft model agency and ask them to help us brush up the prior data.

But the accuracy of the data obtained in this way is relatively poor, and I don't know how long the secondary processing will take.

It is very possible that when the data is processed, Jiancheng has already missed the time to apply for the special class. But as long as the model can be written, it will not be a loss.

There is another way, which is to overthrow the model and start over.

I thought about it carefully just now. In fact, there are still many efficiency optimization points that have been ignored in the model. After solving these points one by one, there will definitely be benefits.

The problem is that if you do this, the underlying logic will also need to be modified, and the amount of work will be relatively large.

Wang Jiancheng sighed helplessly, looked up at the ceiling.

At least it's better than nothing. If there is a problem, we have to find a way to solve it. Let's think about it when we go back.

Chen Nian nodded slightly, but did not speak.

In fact, what Chu Jianshu can do, he can do himself.

But his original intention of joining this project made him unwilling to use this kind of opportunistic method.

This is rather embarrassing.

However, besides this method, is there no other way?

He always felt that he seemed to overlook something. What exactly was it?

On the other side, after hearing Wang Jiancheng's answer, Chu Jianshu nodded in relief.

Fortunately, my first reaction was not to give up, nor to ask myself for help subconsciously, but to plan a solution.

This shows that, at least in terms of xinxing, they have basically passed the test.

So, he patted Chen Nian's shoulder comfortingly and said:

It doesn't matter, that's how scientific research is. Sometimes it takes hundreds or even thousands of failures to get one success.

Take your time, you still have time now.

Understood, Professor Chu. Don't worry, these are minor problems. Let's talk to the teachers of the Mathematics Institute tomorrow to see if we can optimize the algorithm.

It's good to have this idea! Let's go, let's have a late-night snack? I invite you!

Hearing his words, Chen Nian smiled slightly, stood up politely, and did not forget to remind Wang Jiancheng to return the initial settings of the minicomputer.

After tidying up, the three of them walked out of the school gate to have a casual meal. During the process, both Wang Jiancheng and Chen Nian were a little silent.

Of course, Chen Nian was thinking about the fleeting flash of inspiration that flashed in his mind just now. As for Wang Jiancheng, he was probably also thinking about the model.

Seeing the performance of the two, Chu Jianshu didn't say much.

He still thought in his heart that maybe the two children were shocked and didn't want to talk, and of course they might be planning the next step, but no matter what, he shouldn't bother them too much.

Things that should be experienced are always to be experienced. Since you have made up your mind not to interfere, you must carry it out to the end.

However, when he returned to his residence, he still couldn't help feeling a little worried.

Many geniuses actually died in one or several failures.

Especially young geniuses—because their minds are immature at this time, a small setback may cause a big butterfly effect, and eventually cause irreparable impact.

Would Wang Jiancheng be such a person? Is Chen Nianhui such a person?

Chu Jianshu hoped that they were not, but he was not too sure.

——

However, such concerns only lasted less than two hours.

Just after he finished washing, his personal mobile phone received a call from Wang Jiancheng.

On the other end of the phone, Wang Jiancheng said excitedly:

Professor Chu, we want to borrow a minicomputer tonight! We know how to do it!

What? How did you do it?

Chu Jianshu asked in surprise.

It's like this. Senior Chen Nian told me just now. In fact, our thinking at the beginning was wrong.

We've brought the human-to-plane operation into drones!

For us, targeting, then turning, moving, is one continuous movement, but it's not like that!

From the perspective of the machine, the so-called position change is actually the change of one data point.

We don't need to treat it as a continuous function at all, but instead split it into two different models.

The first model, which deals with location selection, is used to determine location targets.

The second model, to deal with policy enforcement.

In this way, the complexity of the continuous function will be greatly reduced, and our training difficulty will be reduced by at least 90%!

Chu Jianshu listened to his words silently, and after a few seconds of silence, he said sincerely:

You guys did a great job!

At this moment, an idea popped up in his mind:

Sure enough, in front of real geniuses, there are really not many frustrations