Imagine that, in the future, we can make decisions on complex problems more quickly and adjust automatically at any time. Many social and industrial problems can also be solved automatically by self learning experience. In the future, front-line rescue workers can analyze Street camera images by image recognition and quickly rescue people missing or kidnapped. In the future, the traffic signal lamp will automatically adjust the time of light change according to the traffic flow, control the time of starting parking so as to reduce traffic congestion. In the future, robots will become more autonomous, and performance efficiency will be significantly improved.
With the increasing demand for collection, analysis and decision from highly dynamic and unstructured natural data, the demand for computation is also beyond the classical CPU and GPU architecture. In order to keep pace with the pace of technological development and promote computing beyond PC and servers, Intel has been researching the special architecture that can speed up the classic computing platform in the past six years. Intel has also recently increased investment and Research on artificial intelligence (AI) and neural mimicry.
Our research in the field of neural mimics is based on decades of research and cooperation. This research was first initiated by Professor Carver Mead of California Institute of Technology. He is famous for its basic work in semiconductor design. The combination of chip expertise, physics, and biology provides a good environment for the creation of new ideas. These ideas are very simple, but they are revolutionary: compare machines with the human brain. The field of research will be highly collaborate and continue to support the further development of science.
As a research topic of Intel Research Institute, Intel developed the first self learning neural mimetic chip named Loihi, which mimics the way that the brain learns how to operate according to various kinds of feedback from the environment. This is a very energy-efficient chip that uses data to learn and extrapolate, and become intelligent with time, and does not need to train in traditional ways. It is calculated in a novel way through an asynchronous pulse.
We believe that artificial intelligence is still in the early stages, and more architecture and methods, such as Loihi, will emerge to improve the standards of artificial intelligence. The inspiration for neural mimicry comes from our current understanding of the structure of the brain and its computing power. The neural network of brain transmits information by impulse, adjusts the weights of synaptic strength or synaptic connections according to the time of these pulses, and stores these changes in synaptic junctions. The cooperative and competitive interaction between the neural network in the brain and the multiple regions in the environment produces intelligent behavior.
Machine learning, such as deep learning, has made great progress recently by using a large number of training data sets to identify objects and events. However, unless these training datasets take into account specific elements, conditions, or environments, these machine learning systems can not be well generalizations.
The potential benefits of independent learning chips are endless. For example, it can provide heartbeat data of a person in various situations -- jogging, eating before or after going to bed -- to a system based on neural mimicry to analyze these data and determine the normal heartbeat under various conditions. The system then continuously monitors the incoming heartbeat data to mark out the situation that is not consistent with the "normal" heartbeat pattern. This system can also provide personalized services for any user.
This type of logic can also be applied to other application scenarios, such as network security. Because the system has learned all kinds of normal mode, it can identify vulnerabilities or hacker attacks when there are anomalies or differences in data streams.
Intel launches Loihi test chips
Loihi studies the test chip, including the digital circuit that mimics the basic mechanism of the brain, so that machine learning is faster and more efficient, and at the same time, the demand for computing power is smaller. The inspiration for neural mimic chip models comes from the way neurons communicate and learn, using pulses and plastic processes that can be adjusted according to time. This will help the computer to implement self - organization and make decisions on the basis of patterns and associations.
The Loihi test chip provides a highly flexible on-chip learning ability and integrates training and inference on a chip. This allows the machine to automate and adjust in real time without waiting for the next update from the cloud. Researchers have confirmed that compared with other typical spiking neural networks, when solving the problem of MNIST digit recognition, to achieve a certain accuracy rate, the learning speed of Loihi chip is increased by 1 million times. Compared with the convolution neural network and the deep learning neural network, the Loihi test chip needs less resources in the same task.
This test chip learning function has great potential and can improve the automotive and industrial applications, including any personal robot in unstructured environment due to the independent operation and application, such as continuous learning, recognition of a car or bicycle exercise.
In addition, the energy efficiency of the Loihi chip has increased by 1000 times compared to the universal computing chip that trains the artificial intelligence system. In the first half of 2018, Intel will share Loihi test chips with famous universities and research institutions to promote artificial intelligence.
Author of this article: Dr. Michael C. Mayberry
Dr. Michael C. Mayberry is the vice president of Intel and the president of the Intel Research Institute, which is responsible for the global research work in the field of computing and communications in Intel. In addition, he also leads the public