MTU-TCP/IP protocol stack-linux kernel-tcp packet loss retransmission-UDP high performance-ai-

http://view.inews.qq.com/a/20161025A0766200
QQ in narrowband era

QQ is a narrow band of the most representative of the product, in that network transmission efficiency is lower than the era, we still remember Google's homepage? Why is the simple page of Google so concise?

Google was born in 1998, is also in the narrow-band era, you will find its home byte size is less than 1024, why less than 1024 bytes, because the Ethernet MTU (that is, the largest transmission unit) is 1024,google in order to let the user in a network packet transmission complete, So it dropped the page size below 1024. This is a performance of extreme transmission.

QQ behind also contains a lot of similar extreme pursuit, the early QQ client installation package is dozens of k size, which is to allow users to download faster, more convenient to use.

QQ Transmission communication is modified the underlying TCP/IP protocol stack, this technology is very difficult, because the need to change the Linux kernel, the reason is because QQ hope that the transmission protocol has TCP packet loss retransmission capability, but also with the high performance of UDP, In this way, QQ can guarantee the transmission of information at the same time can maintain more user network connection, in that era of QQ unit background processing capacity may be other similar products more than 10 times times.

What is AI?

AI is the basic theory, method and technique to study how to use computer hardware and software to simulate some human intelligent behavior. AI includes: computer vision, speech recognition, natural speech processing and other similar human perception, cognitive intelligence.

Now many people think that AI is machine learning, or deep learning. This is very narrow.

Machine learning is the core algorithm of AI. The exact relationship between AI and machine learning is this: AI = a->f (x)->b, machine learning is a fitting function of human intelligence behavior.

The current AI is generally a very elementary stage, although we have some breakthroughs in some "perceptual" intelligence.

For example, we have the ability to face recognition in computer vision, the shorthand ability of speech recognition and other vertical fields have approached or surpassed the average intelligence level of people. But these are very narrow compared to a person's comprehensive intelligence, in essence I think the algorithm itself also needs a higher dimensional breakthrough, rather than simple evolution.

For example, our vision of alexnet to Vgg to the present ResNet, each time the evolution of the algorithm is just stuck in a deeper neural network, more complex neural network connection design, through such a complex neural network learning more sample characteristics to better fit the human brain "perception."

such as "cognitive" intelligent natural language processing, but in the face of this language sequence problem, although our neural network from the early RNN evolution to lstm, and then to the present attention Model, but so far the machine is not fully competent for this complex calculation.

MTU-TCP/IP protocol stack-linux kernel-tcp packet loss retransmission-UDP high performance-ai-