Artificial intelligence — today is the phrase we hear everywhere. He allegedly has in smartphones, there really is some car, probably a steep appliances and certainly hiding somewhere deep inside the robot cleaner or your new mixer.
Actually, of course, all a little different. As usual, you need to start with a story. The first attempts to reflect on the question of intelligent machines date back to the XIX century. Of course, then it had nothing to do with modern concepts of artificial intelligence and, rather, implied familiar to us robots. A more modern concept of AI appeared in the middle of the last century. In fact, then your work entitled “Can a machine think?” wrote Alan Turing is the author of the the Turing test, which was remembered only one of the interviewed employees of our team when Sasha decided to interview all who came to hand.
Briefly, the test is considered completed if communicating with the machine the person is not able to identify a person or artificial intelligence. By the way, in 2014 for the first time in the history of a formally supercomputer has passed the Turing test by convincing 33% of judges that are in front of them. The program, incidentally, was created in Saint-Petersburg. However, later appeared a lot of materials that have questioned this result, and indeed the Turing test is not without flaws. In General, this question finished.
Since the mid-twentieth century and almost to the end of the most ordinary people associate artificial intelligence that accurately captures the essence of both words of the term. That is, with robots or other systems that have real intelligence, but artificially created. This understanding was imposed on science fiction literature and cinema. However, the development of computer technology dramatically changed the meaning given to the term “artificial intelligence”.
Now shove it where popadja. In the beginning of the article, of course, I was joking, but only partly. Producers and developers really began to “reward” artificial intelligence almost every other new device, but the problem is that in most cases we can not verify if there is actually an AI in a particular product or we are openly cheating.
Why is this happening? Partly because the majority of artificial intelligence systems is the cloud services and consumer devices only using this remotely. Another problem is the very essence of modern technology, called AI — they are extremely specialised. For example, the whole “intelligence” of the home surveillance camera is the ability to identify objects captured in the lens of this camera. There are applications that “intelligence” which is limited to only tracking certain features of human use of a device. And so on.
One of the obvious exceptions are voice assistants. In particular, Amazon and Google Alexa Assistant. Cortana and Bixby are lagging behind in their cognitive abilities, and Siri is simply old and limited due to the position of the Apple. A great example of the voice AI of the new generation was Viv. The developers of this assistant at the time, showed a really impressive ability. But then they bought a Samsung, and all these achievements have gone missing. At least Bixby if any of them got very little. So, voice assistants are a good example because they work in one of the most important human spheres — they’re talking. Trying to do this as a man, understanding the context and freely formulated queries. Yes, they also work with cloud services, but there is not stopping.
A good example of the second exception are the technology of AI that does not require connection to the cloud. For the operation of such systems requires a special processor that will work with the systems of artificial intelligence, as an ordinary CPU or GPU (single) for this approach is not the best way.
The first company that introduced single-chip system with a dedicated unit to work with AI, was Huawei. Last year, it announced a SoC Kirin 970 unit NPU (Neural Processing Unit). About this unit not much is known. It consists of 150 million transistors (about 3% of the total number of transistors Kirin 970), and its performance is 1.92 TFLOPS (FP16), but it is for specific tasks, deep learning, machine learning, and other things that falls under the concept of “artificial intelligence”.
Why do we need this block? As already mentioned, in order for the software, with the basis of the principles of AI, worked directly on the smartphone rather than climbed continuously in the cloud. It’s convenient, it’s fast and it’s important for the industry. I note that in addition to Huawei platform with a similar unit, there is only Apple (A11 Bionic) and more recently from MediaTek (Helio P60). It is also rumoured that the unit NPU will get a mainstream chip Kirin 670. If so, then Huawei is the first company who offers relevant especially in devices of the middle class.
So, with the theory clear. And what about the practice? With practice everything is quite simple. First the AI in Huawei Mate 10/10 Pro and P20/P20 Pro is used to operate the camera. Here is one of the most common technologies in this field — object recognition. That is, using artificial intelligence, trying to determine what the object hit the camera lens before the user takes a photo. After that, in accordance with the type of a particular object, the application will set the appropriate settings for brightness, exposure, white balance and saturation. Very good system noticeable when shooting greenery (grass, trees, and so on). Of course, the average user doesn’t care, due to which there is definition, but the world is moving forward, not them.
Also the AI works when shooting in portrait mode, helping to more accurately separate the foreground from the back, and is still involved in the operation of the hybrid Zuma at P20 Pro and when shooting at night. No, the existence of artificial intelligence is not a magic wand that automatically transforms the mediocre camera in excellent. The latest flagship Huawei off great, because the company years this went on. But AI contributes, albeit not the most significant contribution.
And not a camera one. If you believe Huawei technology AI works in some other cases. For example, it is responsible for increasing the performance of the smartphone in General, due to the fact that constantly analyzes the user activity and based on these data, optimizes the system.
Also the AI increases the autonomy of the smartphone, tracking background processes, the user’s habits and optimizing the OS accordingly. In addition, the artificial intelligence involved in the translator Huawei, improving the images that catches the user eye (in a browser or social networks), as well as voice calls, where the AI helps to suppress noise and clear voice. Not all of this can be noticed, even if you know that there is AI, but as in the case of gopher, which is not visible, but it is. And in this case it is a useful gopher.
It should be noted that the question of constantly obuchayutsya on the smartphone. You will not be able to improve some function on your specific device due to its constant use. This is not the case. While the unit NPU (and other similar) need more to speed up the execution of neural network algorithms. Simply put, to accelerate artificial intelligence, although it sounds not really 🙂 the Whole learning process is at the stage of application development.
As always, there may be those who will scream that this is just marketing, and in fact, no separate unit no, but if there is, it’s nothing he’s not accelerating, all is ordinary with the usual algorithms. Indeed, to test the unit NPU directly out of apps Huawei is quite difficult, but possible.
For example, there is a test, as Master Lu, who checks it work hardware solution with three neural network models. It only works with the platforms of Huawei and Qualcomm, but we and that’s enough. So the results below.
As you can see, the block in NPU SoC Kirin 970 is superior to the signal processor Hexagon DSP in Snapdragon 682 835 1.6-4 times and the CPU in Snapdragon 835 and 960 Kirin in these tests, we drop behind much more. By the way, this indicates that the signal processors Hexagon platforms Qualcomm quite a can be used to work with AI. It will be interesting to see how will reveal itself in the Snapdragon Hexagon 685 845.
Compare with Apple A11 Bionic capabilities, unfortunately, no, as there is no multiplatform software for these purposes. But it is the most Huawei.
As you can see, in General, the NPU block really works and works very quickly on the alternatives.
Also, we should remember a video where the smartphone Huawei Mate 10 Pro was driving a Porsche Panamera and was even able to identify the obstacle and avoid it. Moreover, according to Huawei, the process of teaching the AI the new task took only five weeks. Of course, this is highly conditional test, but it shows that to teach the device something brand new it’s possible, would need.
There is another very important point, which is rather important, not specifically now, but in General for the industry. The fact that the NPU unit in the Huawei platform open to third-party developers. Simply put, any developer can create any application that uses the hardware “acceleration” of artificial intelligence. Yes, it is difficult and often very time-consuming and therefore unlikely in the near future we will see a lot of these programs. Moreover, the unit NPU in the case of Huawei there are only a few of the most expensive smartphones. But believe me, it will take only a couple of years, and such blocks are already commonplace, and it means that the developers will make sense to create the software. You will not notice, how the application of AI will be your constant companions.
It is not very clear that with such technologies from Qualcomm, after all its platforms are the most common. But Huawei now ranks third among the largest manufacturers of smartphones in the next year or two is clearly getting out and the second. So the fact that she now has started to actively promote the technology AI in smartphones, is extremely positive. It remains to wait for more available platforms and smartphones with units NPU, as well as the first signs among third-party software.