There is no doubt that deep learning software driving the wave of artificial intelligence. Now, many companies and researchers are all in the effort on the hardware and software for simulating the human brain.

In the aspect of hardware mainly through the study of the simulation of large-scale neural network. GoogleBrain deep learning system such as Google, Microsoft's Adam, etc. But these network needs a large number of traditional computer cluster. For example GoogleBrain has adopted 1000 units each with 16 core processor of the computer, although this architecture showed considerable ability, but energy consumption is still huge.

While IBM from the chip to imitate. : it designs an imitation of the brain called TrueNorth computer chips, and put it as the hardware platform of deep learning. At present, the company is doing research in performance of the chip test.

Depth study the ability strong, the key is one of the (convolutionalneuralnetworks) - convolution neural network algorithm. It contains a large number of nodes (layersofnodes), also called neurons (neurons). Neural network can be filtering through the depth structure like this a lot of information, and then automatically recognize faces or understand the function of different languages.

IBM recently, according to a study in the neural network can support multiple imitation of the human brain hardware, convolution neural network algorithm are also useful.

On July 9, IBM himself a study published in the ProceedingsoftheNationalAcademyofSciences. The study by U.S.D efenseAdvancedResearchProjectsAgency (DARPA), the amount of nearly $1 million. This support is part of the cortex processor project (CorticalProcessorprogram), aimed at promoting can identify the complex situation and adapt to environmental changes in imitation of the brain in artificial intelligence research.

DharmendraModha is IBM's Almaden research center imitation of the human brain computing project chief scientist. He said:

The significance of this research is a milestone, and confirms the concept of an obvious: imitation calculation efficiency of the brain will ascend along with the depth to promote the efficiency of learning, which is a new generation of chips and algorithm paved the way to improve the work efficiency and accuracy.

In 2011, IBM released the first foreign specific TrueNorth and its prototype chip, and developed on the basis of the convolutional neural network deep learning revolution began in 2012. We can see from this, just start is not designed for application of deep learning TrueNorth, instead to have the TrueNorth to promoted the pulse neural network to imitate the real biological neural connections in the brain structure.

Pulse neurons in the neural network is not activated in each iteration spread (and is) in typical multilayer perceptron network, but in the membrane potential reaches a particular value after being activated. This effectively reduces the graphic cognitive or language processing speed.

But deep learning experts think, is also used in machine learning, at least compared with convolution neural network, the pulse of the neural network efficiency is not high enough.

YannLeCun is Facebook pioneers in the field of artificial intelligence research institute directors and deep learning. He has criticized the IBM TrueNorth chip, call it a scientific straw bag. IBM, he says, just copy the appearances of the machine, but there was no deep understanding and the principle behind.

Have objections, but also a voice of support. ZacharyChaseLipton is the university of California (San Diego) team depth study of artificial intelligence researcher. He said,

TrueNorth can promote the realization of neuromorphic computing, you know, deep imitation and understand biological brain is neuromorphic computing.

By comparing the different scholars point of view can be found that deep learning researchers are usually more concerned about is how to make the practical application results to the artificial intelligence technology support services and products.

Also take the old metaphor for example of birds and planes. You may feel convolution neuroscience research bird more; And pulse neuroscience is focusing more on aerodynamics, biology is dispensable.

The benefits of proprietary computer hardware for machine learning more and more obvious. Neuromorphic chip, therefore, did not let a person feel very excited reason mainly is that pulse in the deep learning neural network is not so popular.

Therefore, in order to make the TrueNorth chip can better application of machine learning, IBM must develop a new algorithm to help the convolutional neural network in the form of computer hardware to run on a better.

In the tests, the TrueNorth can be classified to the image data, speed of 1200-1200 frames per second, 25-275 milliwatts of energy consumption. Processor is able to identify model images, these images is 50-100 camera shot at 24 frames per second.

TrueNorth began in-depth study test to get the result, give a person as if it is quite impressive. But everyone still careful is good, Lipton said, after all visual data set in the treatment of 32 x 32 element image, there are a few problems.

But it can affect the Modha. He is still with the feelings of eager to continue the TrueNorth deep learning test. He and his colleagues hope chip test under the environment of unlimited deep learning. This need them in training neural network gradually introduce hardware limitations, rather than from the start to its limit.

Modha also pointed out that the TrueNorth its universality, is one of the advantage of many other deep learning can only run on convolutional neural network hardware, but TrueNorth can accept a variety of types of artificial intelligence network.

"Not only can run TrueNorth convolution neural network (though initially had no intention of the designer), support for multiple include feedback, transverse and feedforward mode of connection, also is applicable to a variety of other algorithms."

Like this kind of biochip only is more superior than other deep learning hardware application can pop up, Lipton said. At the same time, he also suggested that IBM can take advantage of the hardware company, and Google, Intel struggle for special chip design deep learning together.

"I believe the future will certainly be some neuromorphic chip maker, they use their own hardware companies to accelerate the development of the chip to make the industry more focused on the practical application of deep learning, rather than purely biological imitation."