Yang Likun: The Road to Science Reading Notes 2

Let's learn AI series blog: directory index

The low point of artificial intelligence

Last time, I shared the reasons for studying, Yang Likun’s background and his research views on artificial intelligence. This time, I will share how he survived the trough and who can benefit from this revolution.

On December 6, 1987, the two giants of deep learning, Yang Likun and Xinton, met the giants in the laboratory. The former humorous Xinton looked depressed and told a bitter joke: "" Yang Likun couldn't help but said, "Even if 今天是我40岁生日，我的职业生涯也到头了，什么也做不成了。we Belief is firm, and there will be a moment of wavering."

In the 1980s, AI entered its first trough, limited by computing power and training methods, unable to complete large-scale data training tasks. Frustrated, Xinton and Yang Likun didn't know that there is a saying in China that once a child is born twice (I don't know the second birth is not three), at the second trough, Xinton was 60 years old, and Yang Likun was forcibly "graduated" by AT&T.

However, after all, as the philosopher said, " Difficulty is where value lies in." Yang Likun's project name is CIFAR, pronounced "see far", which means "vision". Perhaps it also means that although their research will encounter darkness, dawn will always come.

In the first trough, after several years of groping in the dark, Yang Likun and others proposed the BP algorithm (backward transmission of gradients and adjustment of learning parameters), which made it possible to train AI networks on a large scale and pushed artificial intelligence to the second wave.

In the second trough, what led us out of the dark night was that Sinton and others proposed a deep neural network, tried to make a miracle with great efforts, and piled up the number of network layers into hundreds or thousands of layers (the deepest before was a dozen or so layers). Thinking of being the best in the IMAGENET competition, the accuracy of image recognition has set a record, which is close to the level of human image classification perception. Yang Likun and the others did not expect that there really are: how bold people are and how productive the land is.

In the end, Yang Likun, Hinton, and Bengio shared the top award in the computer industry-the Turing Award in 2019, 32 years after Yang Likun's comrades said the phrase "career is over". His experience tells us that 把你想要写的故事，坚持写下去，终会有答案，even if it may come a little late.

Is artificial intelligence omnipotent?

The two troughs have led us to the current third round of upsurge, which is even called a revolution no less than entering the industrial age, electrical age, and information age. We may be entering the era of intelligence, but it has to be mentioned that even now, artificial intelligence is not intelligent, and sometimes it is even mentally retarded.

He rejects our myth of AI technology, and advocates our understanding of the essence of thinking behind it, rather than terminology formulas.

I think Yang Likun may have suffered from a large number of investors running away halfway. Once the high expectations are not met, the savvy investors will say: Brother Likun, you have good stamina, if you push it up again, we will withdraw first if we have something to do. visible,投资人大多都是好汉——不吃眼前亏那种。

As far as I know, although the current AI progress is in full swing, the core is mainly deep learning, which is better at solving problems, mainly automatic scene problems, such as image recognition, machine translation, speech recognition, etc.

There are still many current AI problems, lack of knowledge reasoning ability, and good explainability (specifically, if you are right, you don’t know why, and if you are wrong, you don’t know how to change it), and the model migration ability is poor, unlike Humanity. According to Yang Likun, in order to be more versatile, the AI ​​model may move in the direction of self-supervision in the future.

Who will benefit from the AI ​​revolution

Some people tell PPT stories with concept hype, get investment from it, and then run away. This is a kind of income. Some people are down-to-earth and rooted in technology, do some practical things, optimize each scene, use AI to reduce manual labor, reduce costs and increase efficiency, which is also a benefit.

For the storyteller, Yang Likun has nothing to say.

For those who do practical things, Yang Likun mentioned a lesson of AI commercialization, which is worthy of vigilance.

The stars of Bell Labs are shining, and they have made many major black technology innovations. However, AT&T has repeatedly missed the commercial application of black technology in its own laboratory, and instead followed it.

This tells us that if managers do not have a deep technical background, they will not have a keen sense of technology. Even if they have created a technological revolution like a digital camera, they will be like Kodak, abandon it, and finally be the technological revolution created by themselves. Killed.

In addition, another inspiration for us is that for followers, 模仿并不是啥问题，Apple also copied Xerox's graphics system in its early years. The key issue is that to pursue the big dream, imitation is only a means, and the ultimate goal is,站在巨人肩膀上，开拓创新。

Back to the question, who will benefit from this AI?

For large companies, looking at the current technology giants: Apple/Google/Facebook/Amazon and domestic Ali, Tencent, Baidu, Byte, Huawei, etc., all of them have already actively deployed AI cutting-edge technologies, and large companies are already the first benefit from it.

For us ordinary people, since the current successful application of AI is mainly classified automation, the value of repeatable and automated work will drop significantly, while industries such as AI无法替代的那些感性的工种内容（文科小姐姐福音）将变得更有价值，service industry, handicraft industry, real estate, and positions such as management, HR, art, etc. worker etc.

epilogue

There is still a long way to go for artificial intelligence to approach human intelligence. At present, it only has data, no reasoning, no emotion, no consciousness, and it is not an intelligent body that can learn by itself. Therefore, there is no need to expect too much.

The free flow of information is the driving force for progress. —— Yang Likun

Tracing back the scientific road of Daniel Yang Likun may give us some enlightenment:

• Practice precedes theory. Do it first if you don't think clearly, the essence of deep learning is a bit of thinking of making up for one's weakness with hard work. Therefore, although deep learning lacks theoretical explainability, it is easy to use. Moreover, it took Shannon's information theory to support the telegraph 50 years after the telegraph was established.
• Study for the knowledge itself, not the diploma. Read widely, form scientific thinking, lay a theoretical foundation, and keep in mind: philosophy is the mother of encyclopedias.
• Be independent, don't forget your original intention, and do things you are interested in. Yang Likun's path to science is itself a declaration of courage. He has been subjected to PUA in the academic world (if it is not for true love, he may not be able to persist until now).

topic exchange

• What problem would you most like AI to help us solve?

In the end, I deeply believe that "serious people get serious results" through reading, thinking, sharing and communicating.

On May 11, 2022