Time Travel: 2014

Before that, Lin Hui responded to various emails as usual.

Although there are a lot of email messages, they are basically recruitment emails from various headhunting companies.

Some time ago, Lin Hui would have clicked through the pages one by one to see how much he was worth in the eyes of various Internet giants?

Now Lin Hui basically doesn’t even look at it anymore.

Lin Hui was really speechless. Why did these headhunters offer me an annual salary that was less than my weekly tax to recruit me to work for you?

Or is this kind of black humor more popular now?

In particular, the HR at the pig farm is even more thoughtless.

Lin Hui remembered that he had clearly replied to this person's email and expressed rejection.

Now I am sending emails one after another...

Uh, I'm speechless.

However, this professionalism is indeed worthy of praise.

Moreover, Lin Hui has always admired serious people.

The level of standing is higher now.

Lin Hui's appreciation was somewhat different from the simple appreciation in the past.

Now Lin Hui is more of a lover of talents.

Lin Hui remembered that the person who sent the email seemed to be the deputy director of human resources.

To be able to climb to such a position, you must have some ability.

Is it too late to keep such talents in the pig farm?

Anyway, Lin Hui didn’t remember anything serious done in the pig farm?

It is not even on the certification list of country M. To put it bluntly, it is dispensable.

Lin Hui rummaged through a pile of emails sent by headhunters/HR for a while.

I found that the wording of the email sent by the HR of Goose Factory was quite good.

So I changed the title and specific treatment of this email.

I sent this recruitment email to the pig farm’s HR and responded.

Although it was an impromptu idea, Lin Hui did not do it as a prank.

At least Lin Hui felt that the salary he offered was much more reliable than the salary offered to him by these HRs.

This kind of thing is just a small episode.

After that, an email from a foreign country in the mailbox caught Lin Hui's attention.

I clicked on the email and saw that the sender was Eve Carly.

Claims to be the leader of the Text Summarization Group of the Natural Language Processing Research Project at MIT.

As if she was afraid that Lin Hui wouldn't believe it, Eve Carly also attached a bunch of certificates to prove her identity in the email.

In fact, Lin Hui remembered the name Eve Carly.

The original owner of the patent "A New Method for Text Judgment, Screening and Comparison" that Lin Hui purchased previously was Eve Kali.

What was she sending the email for?

Could it be that he regretted it?

(⊙﹏⊙), but fortunately Lin Hui did not see Eve mentioning the withdrawal of the patent "A New Method for Text Judgment, Screening and Comparison" in the email.

In the email, Eve highly praised the generative text summarization algorithm developed by Lin Hui in South Wind (Nanfeng APP).

Although Eve Carly's words were full of respect.

People who work on algorithms seem to never hide their respect for the strong.

Although algorithms don’t seem to play a role at all in ordinary people’s lives.

But in fact this type of algorithm is very important.

To some extent, algorithms can even be said to be the core value of applications.

Take the sale of the overseas version of TIKT0K of a certain music in the past as an example.

Companies seeking to acquire TikT0k’s country business considered four options:

The first option is to acquire TikT0k without the algorithm.

But the demand is to speed up the sale while injecting alternatives into the application.

The second option is to slowly transition the algorithm to Country M during a year-long transition period.

The third option is to seek approval from Country Z to sell the algorithm to selected companies in Country M.

The fourth option is for the new buyer to obtain a license from Byte/Section to use the TikT0k algorithm.

Why are the descriptions of these four types of acquisitions different?

In the final analysis, it’s because of the algorithm.

Country M is trying every means to obtain the core algorithm of a certain sound.

The price difference between these different plans may even reach as much as 10 billion US dollars.

From this point of view, algorithms are very valuable in specific situations.

Although the generative text summarization algorithm that Lin Hui had previously developed cannot be compared with the algorithms involved in Xinxin's personalized information push service technology based on data analysis.

But it also has its own unique value.

In the eyes of knowledgeable people, this thing is a treasure house full of treasures.

Anyway, Eve Carly's words are full of praise for Lin Hui.

Eve Carly was curious about how Lin Hui mastered the text summarization technology in Nanfeng APP.

Well, in fact, how Lin Huizainanfeng's generative text summarization processing technology is achieved is very simple.

It only takes about seven or eight steps to easily implement the text summary processing technology in Nanfeng APP:

ⅠBased on deep learning technology, design appropriate model architecture and training strategy.

ⅡDesigning a generative automatic text summarization model

ⅢWith the help of the idea of ​​transfer learning, a generative automatic text summarization algorithm based on the pre-training model is proposed.

ⅣComplete content representation and weight calculation through unsupervised

Ⅴ……

…

The steps are simple to say. Every step is difficult for people in this time and space.

Some ideas are more difficult to think of in this direction.

Some of them are simply impossible to do technically.

Some of them are both unimaginable in terms of ideas and impossible to do technically.

That’s so sad!

For example, the unsupervised training mentioned in step IV.

Now the mainstream research direction has forgotten the direction of unsupervised training.

Be more used to supervised training when it comes to training.

Rather than focusing very much on unsupervised training.

Unsupervised training seems to be a very retro research direction for people in this time and space.

In the eyes of people in this time and space.

Unsupervised training will lead to the phenomenon of divergence of training results, which is not easy to deal with.

The pre-trained model mentioned in step III:

Introducing pre-trained models in natural language processing.

According to the normal timeline, it appeared around 16 years.

This technology is a completely new concept for NLP researchers at this time.

And the deep learning technology in step Ⅰ.

Although people nowadays can think of applying neural network learning technology!

But it cannot do deep learning, although deep learning and neural network learning have similar meanings.

The research on this spatiotemporal neural network is not particularly in-depth.

Although there were neural networks in 2014, the research on neural network learning was not as in-depth as in the following years.

Although people in this time and space have known since 2012 that the deeper the neural network is when learning, the higher the accuracy of the neural network.

But knowing it is not of much use.

Most neural networks in this space-time can only go deep into about ten or twenty layers.

It reaches its limit when it reaches the nearest fifty floors.

It is common for later generations of neural networks to go deep into hundreds of layers and thousands of times.

In a word, people in this time and space have not been able to go that far in neural network learning.

But if you can't go that deep, the accuracy will be easily compromised.

Although Lin Hui is very clear about the path here.

But when the time is not ripe, these things are temporarily not enough for outsiders.