Google artificial intelligence has made new progress only a picture can recognize objects in the pic

Google artificial intelligence has made new progress: just look at a picture can recognize map objects in the image recognition technology of Sina News Beijing standard time on November 4th morning news, computer algorithms usually require tens of thousands of examples to learn a thing, but Google researchers DeepMind has found a way to bypass this process. Most of us have seen something one or two times before we can recognize it. But computer vision recognition and speech recognition algorithms require thousands of examples to be familiar with a new graphic or word. Google DeepMind researchers now find a new way, they made some adjustments to the deep learning algorithm, which only need through an example can recognize objects in the images or other things, they referred to as "single learning". The team tested a large number of tagged images as well as handwritten fonts and languages. The best algorithm can reliably identify objects, but because of the need for huge data, it is very time-consuming and money. For example, if you want to make the algorithm to identify the car on the road, you need to provide tens of thousands of examples in order to achieve reliable accuracy. But it is not practical to collect so much data – for example, it is impossible to provide a long time for a robot to walk in an unfamiliar house. Google researcher DeepMind · Ayers (Oriol Auriol; temperature Vinyals) in deep learning adds a memory component in the system. The team uses a database of labeled images called ImageNet to verify the system’s capabilities. The software still needs to analyze hundreds of images, but then it can learn to recognize a new object with a photo. In essence, it is through the analysis of the unique elements of the picture to complete the recognition task. This algorithm only needs to look at one example, we can achieve the accuracy of the traditional depth learning system. Wen Ya Ersi said, if we can quickly identify a new word, use of this technology will be clearly reflected. This is very useful for Google, so the company can quickly learn the meaning of a new search term. There have been people who have developed a single learning system, but usually not compatible with deep learning system. "I think this is a very interesting approach, which provides a novel way for a large database of single learning." Sang, Wan, director of the brain and Machine Intelligence Laboratory at the Institute of advanced science and technology in South Korea, said: "this has made a technical contribution to the artificial intelligence community, and computer vision researchers may pay much attention to it." However, some people questioned the use of this technology, after all, it is very different from the human learning methods. For example, the Harvard University science department associate professor Sam · Gershman (Sam Gershman) said that human beings typically learn by understanding the image elements, which requires some actual knowledge or attempt. For example, the Segway may seem to self)相关的主题文章:

« »

Comments closed.