What are the gaps between the two AI companies with recognition rates of 97% and 98% in the eyes of Party A?
Editor’s note: This article is from WeChat public account “Fuzhong Finance” (ID: Thecapital), author of the code.
The history of the world economy is a series based on illusions and lies. To gain wealth, the practice is to recognize the illusion, invest in it, and then quit the game before the illusion is recognized by the public.
“I have seen 30 investment institutions this month and will see at least 30 more next month.” On the afternoon of July 2019, Mo Tao told Rongzhong Finance.
Mo Tao is the secretary-general of an artificial intelligence “Maxima” company. He just took office this month. The purpose of recruiting him is also very clear, that is, looking for financing.
“The same investment institution, there will be several FAs to give me docking at the same time, some push the VP, some are partners.” Because there are too many FAs to be docked, Mo Tao can’t even remember these people. first name.
“My attitude is very simple. Whoever leads the person first, who is the list, everything is subject to the final meeting.”
The company that Mo Tao is working on is a AI startup company with a face recognition technology. The founding team is composed of well-known scientists. In 2017 and 2018, it has successively obtained two rounds of A- and B-level venture capital investment, and has also been selected into a certain The list of “Top 20 Chinese Artificial Intelligence Innovation Growth Enterprises” selected by the tripartite service agencies.
“Their financing is not going well,” Liu Wei, an investment banker who helped the company find financing, told Rongzhong Finance.
“The company’s actual revenue in 2018 is 60 million, but the actual arrival is only more than 20 million. Most of the main business is government projects, and the return is very difficult, but the valuation is not cheap.” /p>
“There was a valuation of 2.15 billion before the last round of investment. This round needs to integrate 300 million, 2.6 billion before the investment, and 2.9 billion after the investment. This price will not be picked up this year,” Liu Wei admitted.
“This situation is very common recently. In the past two years, the valuation of too many AI companies was too high, and the bubble could not hold up.”
In fact, Mo Tao’s statement also confirmed this statement.
“Now our attitude is very open, investors think that the valuation is not high, you can first open the price, specifically how you can sit down and talk,” Mo Tao said.
AI revolution and “social people”Counterattack
In 2016, it was called “the first year of artificial intelligence.”
In the spring of this year, a battle between AlphaGo and the world’s top Go player Li Shizhen made the concept of “artificial intelligence” almost overnight. Like science fiction or film descriptions, the first time people who eat melons realize that the fear of being dominated by Skynet is so close to themselves; on the major technology forums, the topics people talk about have become ” The singularity has come” and the three laws of the robot.
However, the so-called “new concept” of artificial intelligence, which has entered the public’s field of vision, has been in existence for more than 50 years.
As early as the 1950s, researchers began to try to give computers “smart” by simulating the human brain.
In their view, the human brain recognizes objects not based on explicit rules, but on intuition. For example, if we see a dog, we know very well that this is a dog, but we can’t say why we know it; in fact, human brain recognition is more like a feature match than accurately defining the characteristics of an object. And this is also the original idea of the “neural network school.”
In the 1970s, computer scientists began to study the feasibility of neural networks in advancing artificial intelligence. However, mainstream academics at the time generally believed that neural networks were mathematically limited and had no future; therefore, the neural network school continued until In the 1990s, it was regarded as a “heterogeneous” in the marginal zone. It is very difficult to get funding and papers.
In 2010, a Chinese computer scientist named Li Feifei from Stanford University organized a machine learning graphics recognition game called ImageNet, which was held annually since 2010.
The interesting thing about this game is that it provides one million pictures per year as a training material for each contestant, each of which is manually marked with objects in the map.
The rules of the game are for players to practice their own programs with this million training pictures, and then let the program recognize some new pictures. Each new picture has a standard answer set in advance, and the entry program can guess five answers, as long as one of the judgments matches the standard answer, it is correct.
In the two years from 2010 to 2011, the error rate of the best scores in the ImageNet competition was over 26%, but by 2012, the error rate dropped to 16%, and it has since fallen.
To 2017, the error rate of machine recognition has droppedIt is 2.3% – this level has exceeded humanity.
So what happened in 2012, so that artificial intelligence technology suddenly appeared a qualitative leap?
The answer is that the “convolution network” was invented.
The winner of the ImageNet contest that year was a research group from the University of Toronto who creatively added several layers of logic between the traditional “input layer” and “output layer” – the so-called ” Convolution layer.”
This research team allows each convolution layer to recognize only one graphics mode of a certain size, and then the latter layer only needs to be identified on the basis of the previous layer; the advantage of this is that each neuron only needs to be processed. A small area of data, and parameters can be reused, which greatly reduces the amount of computation.
In this new model, structures with only one layer of convolution are called simple neural networks (left); those with multiple layers of convolution are called “deep learning” neural networks (right).
This new algorithm is so successful that almost overnight, the neural network represented by deep learning “salted fish turned over and became the master of the house”, from the edge “social people” became the orthodox mainstream – The underlying technical framework of almost all artificial intelligence companies today is inherited from the neural network.
Overall, this is an inspirational story of a “revolutionary” from the edge of society smashing the old world.
And the “convolution network” and “deep learning” turned out to make the new technology shine into reality, human civilization seems to embark on a new evolutionary path.
Redemption: 2018 without progress
This huge “window”, capital must not be missed – starting from 2012, investors from all walks of life began to flock to the AI track.
The first thing to enter is the Internet giant. Companies such as Google and Facebook began to sweep the goods, and they did not hesitate to buy a head scholar in the field of deep learning. For example, the award-winning team from the University of Toronto quickly registered a company. In 2013, it was acquired by Google for $50 million. After that, Google Photos has the ability to search, and then Google can identify the house number of each household from the Street View image taken by the family.
On the other hand, VCPE are not to be outdone, under the hot money, a large number of experts in the field of deep learning began to start their own business with the support of venture capital.