Spread among the widely held such a statement: Google wants to return to China, must rely on AI; and rely on AI back to China, we must engage in a major event. As a result, Li Feifei posted a blog post on Thursday night and made three Twitter launches, announcing an AI product called AutoML.
Overnight, China's technology media exploded and the air seemed to echo that sentence: engage in things, engage in things, engage in things. AutoML is not related to Google's China strategy, and we will not discuss it . Here is to help you find out what is going on in this AutoML.
According to Li Feifei, chief scientist at Google Cloud AI, the goal of AutoML is to lower the threshold for developers, researchers, and business communities to use AI-related tools and frameworks. The popular understanding is probably that this product allows you to train without having to write any code to train an enterprise-class machine learning model. AutoML also means "Automatic Machine Learning".
Surprise? Not unexpected? The legendary machine learning into the annual 500 000 it, say good BAT bee grab AI engineers? Is not feeling just took out the training fee was blown away? More netizens said with astonishment: "Is not to say that together with AI to leather a lot of work life? Why did I work hard to learn AI, the result was first revolution? & rdquo;
Of course, the reality is not that cruel. But Google's moves are not isolated cases, hidden behind the scenes, has been repeatedly mentioned as "AI democratization" and also reflects "AI technology we learn today may be useless."
Funny, right? The original AI revolution was the AI engineer?
Google announced this let yan nong haunting thing, called AutoML Vision, is the first product of the entire AutoML system, focusing on automatic production of image recognition in the field of model.
We describe in a straightforward way how this system works:
If I wanted to do an AI system capable of image recognition before, then I need to set up the training process in the development framework, complete various training and deployment, and import the data set. The whole process needs to be programmed.
But with AutoML Vision, I did not have to write a single line of code. Just follow the instructions and drag the picture I wanted for training onto the system, and wait patiently. A well-trained machine learning model was baked.
For example, if you want to train a model to identify if your cat owner is upset (how bored ...) then just put your cat in AutoML Vision Photos when it's angry, photos when you're happy, etc. Then you get a recognition program. Use it even on the phone camera, you can let the AI to understand the cat adult emotions.
Is not quite good?
AutoML Vision drag and drop image interface
Behind this is Google's use of transfer learning in deep learning. The training process that accumulated when Google trained the image recognition model was migrated to AutoML, thus saving the subsequent similar model development process.
In simple terms, AutoML is a bit like a Google cloud to build a "solution formula". After the candidate does not need to know how the formula is coming, just need to put the problem into the can get the answer. Of course, this is only a brief account of how it works, in fact not so easy. Especially in the debugging process, the compatibility of different model requirements and systems is a big problem.
All in all, this product, along with the thinking behind it, is a boon to companies that want to do machine learning and lack expertise and talent. What it canceled was the process of building machine learning models through code, as well as complex debugging. Only for farmers who have entered the specific data to enter this operation. Greatly reduce the workload of machine learning training, especially programming.
But do not be too optimistic. Although AutoML has not yet officially released, the real effect needs to be considered, the use price is unknown. However, according to the current information, the amount of data required by AutoML to generate a customized model is still very large, not a non-foundation developer can handle it.
And it can only accomplish relatively simple tasks, and can only apply the training program given by Google. If you want to create a more complex machine learning system, using a unique algorithm for training, then programming is inevitable.
So, now look really hard and strive to enter the field of AI development friends can rest assured. In addition to image recognition, Google plans to expand AutoML services to include translation, video and natural language processing in the future. This may mean that the primary AI program is automatically generated, quickly copied to various industries is not far off.
Although Google said AutoML is the only such product yet, businesses are deploying similar businesses. Such as Amazon's Amazon SageMaker, and Microsoft's yet-to-be-released custom image recognition model service. Including the domestic Baidu, also in its AI open platform launched a customized image service.
The reason why Google "engaged in a major event" is mainly due to the higher degree of automation of AutoML, especially to solve the auto-build training model and transfer these two major issues.
However, behind Google so hard, it seems that there are five characters written: AI democratization.
Democratization of AI, who has to cancel the centralization?
Last year in March, just joined Google shortly Fei said that the next step in artificial intelligence is to complete "AI democratization". After releasing AutoML this time, she once again said that due to the scarcity of resources, most enterprises can not develop personalized models, so AutoML appeared to further promote AI democratization.
Then the question is: AI democratization, in the end is for whom the dictatorship? Who to cancel the centralized?
Someone said that most of the advanced AI technology is now mastered in the hands of several large companies. Of course, AI democratization should make everyone the owner of the AI and break the dictatorship of the giant.
I can only say that you are stupid ah?
Is Google would have great effort, in order to disintegrate their hegemony? of course not. Like AutoML's product ideas show, omitted the developer's technology threshold, Google suffered a loss? No. Google has more users, its own algorithmic advantages have virtually been expanded. And AutoML user training model is to be deployed directly in the Google cloud, apparently it is a bundle of disguise, hoping to barbed AWS from the perspective of trickery.
For the smaller companies and individual developers most desperate to get "democracy," the giants 'touts of developers' empowerment and go-to-technology thresholds are "democratically" by no means a gratuitous gift, in exchange for small Developers are closely dependent on the ecology of the relationship. Really demolished by the democratization of AI, in fact, is sandwiched between large companies and small developers between middle-level companies, or algorithm companies, technology companies.
Currently in the AI market around the world, such medium-sized companies are ubiquitous. Of course, its existence is meaningful and valuable. For the giants, the technical ability to enter all walks of life, the development of a wide range of applications is completely impossible, then it depends on the developer to do these things, their own platform to do the service and technical ability of the output like It's
But can small teams and individual developers do these in-depth industry or creative development? In fact, it does not work, because from scratch technology is too complicated, the requirements of the personnel is too high, the average developer and small businesses simply can not afford to play.
So there a large number of mezzanine companies, their technical barriers as the main industry support. Use heavy technology and talent to do little apps that actually do not look "heavy" or charge third parties for developing custom AI models.
The AI democratization advocated by giants such as Google and Intel is in fact putting the work that needs a great deal of "AI workforce" into the automation model. Back up the underlying technology development rights, down directly touch the subdivision development scene.
To put it plainly, small companies, individual entrepreneurs, thin-throated areas want to introduce AI, you must be able to cross the cost of technology companies to play on their own AI. The AI democratization of the giant, of course, is the hope of disrupting businesses that rely on "secondhand technology" and "hiring power" cards to give their share to smaller developers. Allows developers to provide only the creative and operational, marketing and other capabilities to activate the broad prospects for AI.
So the so-called AI democratization, the real collapse is AI non-creative work to create barriers to entry.
After the automation strategy represented by AutoML is further implemented, the most likely impact is the deepening of the AI industry chain. Some seemingly large companies, but only created a small value, just rely on AI technical barriers to occupy the market share and capital concerns, probably the "Democracy Movement" in the most dangerous party.
Today's AI, will not run the horse's car?
And finally, we may open another hole: AI democratization, though it sounds beautiful, is not necessarily all good news.
Needless to say, the progress of AI democratization brought by such products 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 find that AI's industrial chain and development rules have changed dramatically.
But is this a good thing for businesses that are on the edge of the industry? As mentioned at the beginning of the article, AutoML first released, many programmer friends first reaction is to ridicule their AI white school. But you know, how long they can learn. So for those who want to enter the field of business, a similar fear will be deeper?
We know that in the vast majority of consensus, the core of machine-oriented AI revival today is composed of three factors: algorithms, calculations, and data. Interestingly, all three of these factors are now fast "being democratized".
Algorithms, AI development framework began to become more intelligent, AutoML such products can even be programmed automatically to achieve algorithmic tools and modular. Calculating the power, all kinds of new hardware and processing come, the cost of computing almost every day in the fall. Data, more and more free data sharing has become the norm. And the cutting-edge AI technology debate has begun to question the necessity of large amounts of data for deep learning.
All in all, AI's threshold is lowered almost once in a while. But from a business and developer perspective, the situation is a bit distressed: Will the AI capabilities now deployed be phased out in a month?
We know that the car is the beginning of the carriage can not run. Although it turns out that cars are definitely stronger than cars and buying cars is right, everyone does not want to be a passenger of that super classic car. Not to mention AI this thing, even classic cars that collect value is not.
At this moment, I do not know if there is an entrepreneur who has just set up a team of luxury machine learning experts. Suddenly he found out what he really wanted to do. Just look for a new product under his own control and wait for the day. .....
Perhaps today, companies are not reaping the benefits of a gradual decline in the AI development threshold, but rather panic in the face of rapid iteration of the AI system. After all, we do not know whether the things we do today will prove meaningless in a few months.
Especially for Chinese companies, it is customary to follow the trend of the mouth, the wind is on the wind is small. AI is so complex and changing every day, who has the mood to hide and seek with it ah?