“What artificial intelligence cross talk, is simply an artificial retarded. — Guo Degang never said that.

01

A week ago, in a variety show, Dr. Li of Jiaotong University angrily criticized Guo for not knowing crosstalk, and proposed his own formula crosstalk theory. Where to put the punch line and when and how long the audience should laugh are calculated by the finite element formula of crosstalk.


After watching the show, I realized that if I could generalize this theory, I could use linear algebra to write lyrics, calculus to compose music, aerodynamics to write love poems and so on. And the ultimate version of this formula cross talk should be artificial intelligence cross talk, after all, now all kinds of alpha dogs flying, Siri Xiaona accompany me to sleep all night, ask an AI when cross talk funny man isn’t so easy?

My own thinking: although three points are funny and seven points are flattering, after all, the straight man’s workload is not much, and the technical difficulty seems not high. Use the word “um” proficiently. “, “Ah.” “And” Me?” “That’s it!” “Go to hell with you!” Five phrases can easily handle 80% of a conversation. So I aspire to the training of an artificial intelligence field of Uncle Yu Qian.

02

The AI Yu Qian uncle’s code is impossible to write from scratch, because they will not. My knowledge of GAN (generative adversarial network) and RNN (recurrent neural network) is limited to concepts. So go to GitHub and find an open source project that uses GAN for conversation training [1].

In a nutshell, the idea behind this conversational system is to build a generator and a discriminator, respectively. The generator can be used to generate dialogue imitating Yu Qian. The discriminator is used to judge which sentence is really said by Uncle Yu Qian and which paragraph is generated by the model. The training process is the generator to generate dialogue, discriminator to pick out the truth of the process.

With a ready-made network model, my main task was to collect training data, which was the text of Guo Degang and Yu Qian’s crosstalk.

So I wrote a little crawler, went to the Internet to climb down nearly one hundred relevant texts, after screening processing, get fifty guo Degang Yu Qian cross-talk. This is basically the only data set in machine learning that looks like it has sound.


With the data set and code in hand, you’re ready to start training. Here I would like to thank Yao Xiaozhe for lending me the cloud server, so that I can use the high-performance graphics card to let AI Yu Qian speed in the electronic world.

After a day or two of flying through the electronic world, I came up with a rough model of Mr. Yu. Before testing this model, I thought I trained out of the AI community yu Qian, the results found that the artificial intellectual retardation is inferior to the level of small fart.

03

Depending on whether it is their own hands-on training out of the model, or to see how AI Yu Qian’s performance.

Can’t wait to let Uncle Yu Qian chua’s “red high heels” when the straight back.


This “um”, “ah”, “ao” three words I can still hear the voice of uncle Yu.

Say something to Calorie 101, by the way.


Sometimes this AI can also be true · modest elder uncle attached, seems more normal.


And sometimes become real · artificial retarded, and is the kind of dead deny.


And most of the time it’s the wrong medicine.


04

After all, Dr. Li is a doctor with the honeyed confidence and courage to anger Guo degang. And the ARTIFICIAL intelligence made by a man who doesn’t even have a master’s degree is a fine artificial retarded.

To reflect on why the model’s performance is so alarming, there are three main reasons:

1. The quality and quantity of training data sets are not high enough.

2. My lack of understanding of the code (obviously I directly used others’ code) leads to my inability to accurately judge the current state in the training process. As the saying goes, IT is called blind XX training.

I’m not a good funny, unable to inspire AI Yu Qian all his strength.

Compared with my unreliable AI Yu Qian, Yao Xiaozhe’s computer vision project CVTron[2] can be said to be quite reliable. Yes, this is the part about thanking the sponsors and helping with the advertising.

Accompany AI Yu Qian uncle sang a night “red high heels”, “blank grid”, “calories” and “reserved”, talked about one hundred times in the family prominent family, about fifty fart with Yu Qian’s relationship. AI Yu Qian or a pair of take the wrong medicine.

In line with the purpose of serving the society, open source a wave of crosstalk data set [3], others may come up with a super powerful AI Yu Qian.

[1] github.com/MWransky/di…

[2] github.com/unarxiv/cvt…

[3] github.com/unarxiv/cro…