The saying goes, “Google wants to go back to China and it must rely on AI; when it comes to AI back to China, it must do something big. As a result, Li Feifei released a blog post on Thursday night and then posted three Twitters to announce an AI product called AutoML.
Overnight, the Chinese science and technology media exploded, and the air seemed to echo the phrase: Doing things, doing things, and doing things... Is AutoML related to Google’s China strategy? We will not discuss it. What we want to help you find out here is what AutoML is going to do.
According to Li Feifei, chief scientist of the Google Cloud AI project, the goal of AutoML is to lower the barriers for developers, researchers, and enterprise groups to use artificial intelligence related tools and frameworks. The general understanding is probably that this product allows you to train an enterprise-level machine learning model without writing any code. AutoML is also the meaning of "automatic machine learning."
Amazed not surprised? Not surprising? According to legend, the Institute has learned 500,000 in machine learning every year. Say good BAT is mad at AI engineers? Did you feel that the training fees you just left were blown away? More netizens were surprised to comment: "Isn't it meant that I used AI to make a lot of work together? How did I work hard to learn AI and the result was revolution?"
Of course, the reality is not so cruel. However, Google's actions are not alone. The hidden behind it is the "democratization of AI" that has been repeatedly mentioned, and it also reflects that "the AI technology we learn today may be useless."
Funny? Is the original AI revolution the first AI engineer?
Google's announcement of what has become a blizzard in the code country is called AutoML Vision. It is the first product of the entire AutoML system and focuses on the automatic production of models in the field of image recognition.
We describe how this system works in an easy-to-understand way:
If I wanted to do an AI system that can perform image recognition before, then I need to build my own training process on the development framework, complete various training deployments, import data sets, and the whole process needs to be done programmatically.
But on AutoML Vision, I don't have to write a single line of code, just follow the instructions, drag the pictures I hope to train into the system, and wait patiently, a trained machine learning model will come out.
For example, if you want to train a model to identify whether your cat owner is angry (how boring it is...) then just put your cat in AutoML Vision. Photos, photos when it's angry, photos when you're happy, etc. Then you'll get a recognition program. Use it to take a photo with your mobile phone, you can let AI understand the happiness and joy of cat adults.
Is it quite God?
Behind this is that Google has used Transfer Learning technology in the field of deep learning. The previous training process of Google trained in image recognition model was migrated to AutoML, which saved the development process of subsequent similar models.
In simple terms, AutoML is a bit like a “solving formula” built by Google Cloud. After that, the candidate does not need to know how the formula is coming from, and only need to put the problem in order to get the answer. Of course, this is only a brief explanation of how it works. It is actually not that easy. Especially in the debugging process, compatibility of different model requirements and systems is a big problem.
All in all, this product and the ideas behind it are indeed a boon for companies that want machine learning and lack professional skills and talent. What it cancels is the process of building a machine learning model through code, as well as the complex debugging work. For the coders only the operation of entering specific data is reserved. The workload in machine learning training has been greatly reduced, especially in programming.
But don't be too optimistic. Although AutoML has not yet been officially released, the actual effect is still to be considered, and the use price is unknown. However, judging from the current information, the amount of data needed by AutoML to generate a customized model is still very large, and it is not a basic developer's ability to handle it.
And it can only accomplish relatively simple tasks, and can only apply Google's training program. If you want to make a more complex machine learning system and use a unique algorithm for training, then programming is inevitable.
So, at this point, we can rest assured that our friends who are truly attentive and are striving to enter the field of AI development. In addition to image recognition, Google plans to expand AutoML services in the areas of translation, video, and natural language processing. This may mean that the primary AI program is automatically generated and quickly copied to various industries.
Although Google said that AutoML is the only one of its kind at the moment, in fact, everyone is deploying similar services. Amazon's Amazon SageMaker, as well as custom image recognition model services not yet officially released by Microsoft. Including domestic Baidu, it also launched customized image services in its AI open platform.
The reason why Google said this time "has a big event" is mainly because AutoML has a higher degree of automation, especially solving the two major problems of automatically building training models and adjusting parameters.
But behind Google’s hard work, it seems that there are five big characters: AI democratization.
Democratization of AI. Whose centralized power should be removed?
In March of last year, Li Feifei, who had just joined Google, said that the next step in artificial intelligence is to complete "the democratization of AI." After this release of AutoML, she once again stated that due to the scarcity of resources, most companies are unable to develop a personalized model, so the emergence of AutoML is to further promote the democratization of AI.
Then the question arises: Who is the dictatorship for the democratization of AI? Want to cancel who's centralized?
Some people say that most of the advanced AI technology is now in the hands of several big companies. The democratization of AI is, of course, to make people become masters of AI and crack giant dictatorship.
I can only say that you are silly?
Do Google spend a great deal of effort to disintegrate its hegemony? of course not. Just as AutoML's product ideas show, omitting the developer's technical barriers, did Google lose? No. Google has more users, and its own algorithmic advantages have virtually expanded. And AutoML's user training model is to be deployed directly on the Google Cloud. Obviously this is also a disguised bundle. It is hoped that the AWS will be able to penetrate from the point of view.
For the small companies and individual developers most eager to get "democratic", the giants peddling the developer's empowerment and deterring the technical threshold of "democracy" are by no means a benevolent gift, in exchange for small developers. Dependency on the ecology. The real democratization of AI is actually a middle-tier company sandwiched between a large company and a small developer, or an algorithm company or a technology company.
Currently, such medium-sized companies are ubiquitous in AI markets around the world. Of course, its existence is meaningful and valuable. For the giants, it is completely impossible to break the technological capabilities into all walks of life and develop a variety of applications. Then it depends on the developers to do these things. It is good to be the exporter of platform services and technical capabilities. Now.
But can small teams and individual developers do these deep industries or creative development? Actually, this is not enough, because if you start from scratch, the technology is too complicated and the requirements for talents are too high. Generally, developers and small businesses cannot afford it at all.
As a result, a large number of mezzanine companies emerged, and they used technical barriers as the main industry support point. Use heavy technology and talent to do small applications that don't seem to be "heavy," or to charge third parties in ways that develop customized AI models.
The democratization of AI advocated by giants such as Google and Intel is actually putting the work that needs a lot of "AI labor force" into the automation model. The underlying technology development rights are withdrawn upwards and down to the subdivision development scenario.
To put it bluntly, small companies, individual entrepreneurs, and talented people who want to introduce AI must be able to cross the cost of technology companies and be able to afford AI. The democratization of giants' AI is, of course, the desire to dismantle those companies that rely on "second-hand technology" and "employment ability" cards to split their share among smaller developers. Let developers only provide creative and operational, market and other capabilities to activate the broad prospects of AI.
Therefore, the so-called democratization of AI has actually disintegrated the entry barriers created by non-creative work in AI.
After the implementation of the automation strategy represented by AutoML, the most likely impact is the de-levelization of the AI industry chain. Some seemingly large companies have only created very little value, relying solely on AI technical barriers to occupy market share and capital concerns, probably the most dangerous party in this "democratic movement."
Today's AI, will it be a horse that runs a horse?
In the end, we may also want to open another brain hole: Although Democratization of AI sounds beautiful, it may not be all good news.
Needless to say, the progress of AI democratization brought by products such as AutoML can be said to be visible to the naked eye. If you look back one or two years ago, it is easy to see that the AI's industrial chain and development rules have changed dramatically.
But must this be a good thing for companies that are not on the edge of the industry? As mentioned at the beginning of the article, when AutoML was released, the first reaction of many programmer friends was to ridicule their AI Bai Xue. But you have to know how long it took them to learn. So for those companies that want to enter this field, will similar fears be deeper?
We know that in most consensuses, today's machine learning-led revival of AI consists of three elements: algorithms, computational power, and data. Interestingly, these three factors are now rapidly "democratic."
Algorithmically, the AI development framework is becoming more and more intelligent, and products such as AutoML can even be automatically programmed to implement algorithmic tools and modularization. In terms of computing power, all kinds of new hardware and processing have arrived, and the cost of computing has fallen almost every day. On the data, more and more free data sharing has become the norm. And in the frontier AI technology debate, it has begun to question the necessity of a large amount of data for deep learning.
In short, almost every time, the threshold of AI will be lowered once. But from the perspective of companies and developers, this situation is a bit distressing: Will the deployed AI capabilities be eliminated in a month or so?
We know that cars can't run without horses at first. Although it was finally proved that the car was absolutely stronger than the carriage and it was right to buy a car, everyone did not want to be the passenger of that super-class car. Not to mention the AI thing, not even the collection value of a classic car.
Just like this moment, I don't know if there is an entrepreneur who has just set up a team of luxury machine learning experts. Suddenly he found that he wanted to do what he wanted to do. He just took Google's new product for personal use and waited until one day to get it done. .....
Perhaps today, what companies are gaining is not the benefits of the gradual reduction of the AI development threshold, but the panic that arises from the rapid iteration of the AI system. After all, we do not know whether the things we do today will prove meaningless after a few months.
Especially for Chinese companies, they are accustomed to catching the tuyere. When the wind comes, the wind is small. The AI that is so complex and changes every day, who has the feeling of hiding with it?
You may have different opinions and opinions. Welcome attention and comments.
Attention address : https://steemit.com/@niubishini111