March 27 local time, the American Computer Society announced that the 2018 Turing Award will be awarded to three pioneers in the field of deep learning-Joshua Yas Benguio, Jeffrey Sinton and Jarn Le Quin, to praise them for promoting deep neural networks as an important part of computer technology. The Turing prize is known as the “Nobel Prize in the computer World,” and the three winners are also famous scientists in the field of artificial intelligence.
Benguio is a professor at the University of Montreal, Hinton is vice president of Google, honorary professor at the University of Toronto, and Le Quin is a professor at New York University and chief artificial intelligence scientist at Facebook. “Artificial intelligence is currently one of the fastest growing disciplines in all fields of science and one of the most hotly debated topics in today’s society. Chery Pankaik, president of the American Computer Society, said this was due in large part to the remarkable progress made in recent years in the area of deep learning, which was based on the foundations of Benguio, Hinton and Le Quin.
Expand the single-layer neural network into multiple layers
To explain the contribution of three scientists, we must first talk about artificial neural networks. The so-called artificial neural network, refers to the imitation of human neural mechanism, in the computer to simulate a layer or layers of cells known as “neurons”, so that they interact with each other through weighted connections.
By changing the weighted values of these nodes, the computational performance of artificial neural networks can be changed.
Benguio, Hinton and Le Quin recognize the importance of forming a more “deep” artificial neural network by building multiple layers of neurons, which is one of the origins of the word “deep learning”. “Three winners can indeed be called founders in the field of deep learning.
“The early realization of neural networks is single-layer, they expand single-layer neural networks into multiple layers and put them into application, and they have achieved good results in many tasks, such as image recognition, speech recognition and machine translation,” Zongchengqing, a researcher at the State Key laboratory of automation at the Chinese Academy of Sciences, told the science and Technology daily. “By greatly improving the ability of computers to understand the world, deep neural networks are changing not only the field of computers, but also every area involved in scientific and human behavior.
“Jeff Dean, senior vice president of Google, said.
Never recognized for sticking to an industrial outbreak. In the 80, scientists began using artificial neural network models to help computers complete pattern recognition tasks and simulate the intelligence of the human brain.
Hinton, Benguio and Le Quin have been sticking to this line of thinking until this century, although at first their ideas were not recognized. “The computer science community has come to realize that this approach is not bizarre, which is a good thing.
In an interview with the BBC, Hinton said that over the years everyone felt that artificial neural networks were not worth mentioning. Hinton, who has championed machine learning methods since the early 80, and other scientists have proposed using “artificial neural networks” as the cornerstone of machine learning research.
Now, deep learning has been widely adopted in the field of artificial intelligence. “A large part of the reason is that there has been a fundamental improvement in computer performance. Zongchengqing told the science and technology daily that there was a wave of artificial intelligence in the 890, but the storage capacity and computing power of computers were very limited at that time, and only single-layer neural networks could be calculated.
Today, computer performance is greatly improved, coupled with a large amount of data support, in the multi-layer neural network training large-scale data can be quickly realized.
Deep learning still has to make a bigger breakthrough “The application of deep learning technology in computer field is very common at present. From an application perspective, this approach does enable the processing of many tasks to achieve optimal results.
But in Zongchengqing view, the deep learning technology, which has been widely adopted, still needs a bigger breakthrough in the future. Aside from its progress, deep learning techniques at this stage will not allow computers to understand language, voice, and images in depth like people. Smartphone assistants, for example, look articulate, but don’t really understand what we’re saying.
If you say to it, “stinky tofu is so fragrant,” its response will be inexplicable, or give a good answer to how to understand. Zongchengqing believes that, taking natural language understanding as an example, the next step is to allow the machine to reason and calculate semantics, concepts, and not only stay at the signal level of processing.
This involves a lot of problems with the combination of neuroscience, cognitive science and computational Sciences. At the same time, artificial neural network is still a “black box”, the interpretation is relatively poor. Let it translate English into Chinese, if there is an error, which link causes the error is still difficult to explain.
Dealing with a particular task requires several layers of neural network to achieve the best, and there is no reasonable explanation, can only rely on experience and experiments to test a lot. “Deep learning technology will gradually mature after a period of development, into a relatively stable platform period.” Further breakthrough requires a great improvement of artificial neural networks, or the proposed new models and methods based on multidisciplinary cross-research, including brain science.