Go homepage(回首页) Upload pictures (上传图片) Write articles (发文字帖)
The author:(作者)aaapublished in(发表于) 2014/7/27 17:52:33 Microsoft AI Adam beat Google’s Brain，
Microsoft artificial intelligence system Brain-Adam Adam beat Google, Microsoft's artificial intelligence system, Brain, Google AI, AI-IT information Microsoft AI Adam beat Google Brain
Summary: Adam compared to previous systems, image recognition accuracy by a factor of 1 time.
We are entering a new era-the era of artificial intelligence. Relying on a lot of experiments, Google, Facebook, big technology or Internet companies such as Microsoft and Apple are seeking more powerful artificial intelligence technology, including speech recognition, language translation, computer vision, and image recognition.
Beat Google Brain
In the field of artificial intelligence, compared to Google's opponent is in the forefront. Recently, the company acquired the startups in a neural network, University of Toronto DNNResearch. It is reported that the company founder of the machine "deep learning" (Deep Learning) leaders. In addition, Google has openly discussed a new artificial intelligence technology, which relates to the Android voice search on your mobile phone. Google in terms of accuracy of speech recognition and computer vision has set several records.
But now, Microsoft's research team said their newly developed a new artificial intelligence system named Adam. Microsoft is also at its recent Summit of the Organization of academic was mentioned for the first time this new technology. Micro- soft, said Adam compared to previous systems, image recognition accuracy by a factor of 1 time, particularly with respect to identifying dog breeds or vegetable species. "Adam is designed to explore the mysteries of the human brain. "Microsoft Research Director Peter Lee (Peter Lee).
Peter Lee said ImageNet 22K standard test, the Adam of the neural network in image recognition faster than Google's Brain. ImageNet 22K standard mainly to Google's online services (including Android devices, such as speech recognition and Google maps) provides an intelligent evaluation of calculation method. This test requires the system to deal with 22,000 pictures data capacity. Before Adam, and only a handful of models of artificial intelligence to be able to deal with the vast scale of data entry, which includes Google's Brain.
However, Adam was not by virtue of "deep learning" algorithms to beat Google, it's the way to win is to optimize the equipment data processing and adjustment systems and the communication between devices. Microsoft is the one of the founders of the Trishul Chilimbi. Funny human being, Trishul Chilimbi had no expert in the field of artificial intelligence, but rather engaged in large-scale computer systems research.
Adam works
With other "deep learning" system similar to Adam runs on a set of standard computer server array, in fact it uses Microsoft's Azure cloud computing service. "Deep learning" through building the neural networks of the brain system that more closely mimic the human brain works. Adam on servers running process needs a lot, just as in the operation of human brain in complex neural networks as you want. The difference between the two is that Adma is the use of "asynchronous" (asynchrony) technology.
As computing systems become increasingly complex, to different parts of the system becomes more and more difficulties in communication with each other. However, "non-synchronous" technology can help solve this problem. Fundamentally, the "synchronization" system is divided into more than one small portion of their "share" their operations and integrated into the former, are run independently of each other. The problem is that although the technology to good use on Smartphones and laptops (internal calculations for smartphones and laptops scattered throughout many different chips), but contains a lot of different server systems (such as neural networks), the performance of this technology is not successful.
However, many researchers and technology companies, including Google in the past several years have put a lot of effort in this area, it is by virtue of the University of Washington and Microsoft Research's "HOGWILD!" Edge technology.
"HOGWILD!" Technology initially within each processor can run independent of the machine. Different chips and even write to the same storage unit, and each other will not be "overlapping" effect. For most systems, this technology is by no means what "good stuff", they may result in data conflicts, because a device covering the device has completed the data encoding. However, in some cases, this technique works well. As University of Washington researcher shows that "HOGWILD!" Can dramatically improve the speed of a single device. Microsoft Adam on this basis is improved using "HOGWILD!" Technology. "We're crazier than HOGWILD!, because Adam asynchronous stronger. "Chilimbi said.
Despite the distribution of nerve network extremely dense and data conflicts is risky, but it works, because if the system is to avoid data conflicts, this "conflict" will bring the same kind of operation.
Microsoft says that Adam can help neural networks more quickly and more accurately to "train" the cognitive abilities of the device on the picture. What's surprising is, Adam runs turned out to be the traditional computer chips, rather than a dedicated GPU for graphics processing. At present, a lot of "deep learning" is used by the system GPU, in order to avoid data bottlenecks. But Chilimbi, said Microsoft's Adam chose different paths.
Neural network relies on a huge amount of data, what data it is processing significantly more than standard computer CPU chips. This is the reason why they need to be spread out over many machines. As an alternative, GPU can handle so much data, and faster. But the question is, if artificial intelligence data has not been fully run in multiple GPU with GPU or matched to a single server, the system would collapse.
In addition, the speed of communication system in the data center was not enough to keep up with the speed of the GPU to process information, which will cause the data plug. Because of this, some experts say, GPU is not the best choice for handling large neural network system.
Microsoft has described Adam as a "fantasy" system, but some deep learning experts in the field say Microsoft Adam system works with Google is not entirely without artificial intelligence. Experts say that if you do not know more details about how to optimize Microsoft Neural network, it is difficult to know how Microsoft Adam team to achieve the results they claimed.
Former Google Brain Department researcher mate·saile (Matt Zeiler), "said Microsoft research with human beings have been found contrary to the research success of, but this is the highlight of it. "He said that if Microsoft is able to add more devices, Adam that the higher your recognition accuracy will be.
Future potential
Microsoft said Leopold Lee, Adam is still at an early stage of the project. Up to now, it is merely an application for testing. Test results show that, when you take pictures with a cell phone, Adam can accurately identify objects on the picture. Pete · Lee uses Adam to identify dog breeds and poisonous insects. Microsoft has not disclosed when applied to a mass market launch of the plan. However, Leopold Lee said that Adam can be certain of in e-commerce in the future, robots, as well as the sensitivity analysis has been "reused". Microsoft is also being discussed, if Adam is running on programmed gate arrays (field-programmable arrays) or running a conventional software on a processor, and its efficiency will improve. In fact, Microsoft has tested this type of chips to boost Bing services.
In addition, Leopold Lee firmly believes that Adam will be able to become a true artificial intelligence, intelligent technology with human beings process language, Visual and text content are nearly. But, want to achieve this level, still needs to do a lot. Since the 50 's of last century, mankind had already begun working towards this goal. But now, you can be sure of is, we're a step closer.
赞