Science and Technology Daily, Beijing, July 30, according to a recent report by the physicist organization network, researchers at the National Institute of Standards and Technology (NIST) have developed a silicon chip that accurately separates the ability to emit signals for future neural networks. Research provides a potential design method.
The human brain has billions of neurons (neural cells), and there are thousands of junctions between each neuron. Many research projects are dedicated to making artificial neural network circuits to simulate the brain, but traditions like semiconductor circuits Electronic devices, often unable to meet the extremely complex line requirements of a properly functioning neural network.
The NIST team recommends using light to replace current as a signal medium. Neural networks have demonstrated superior capabilities in solving complex problems, such as quickly identifying pattern types and accurately analyzing data. Light applications will further accelerate signal propagation and eliminate Charge interference.
NIST team physicist Jeff Chilis said: 'The advantage of light is that it can further optimize the performance of the neural network, enabling it to perform accurate scientific data analysis, such as searching for terrestrial planets and for quantum information science. And accelerate the development of high-intelligent driverless vehicle control systems. '
According to reports, NIST designed chips through the vertical stacking of two layers of photonic waveguides, overcome the main problem in the application of optical signals. This structure limits the light to a narrow route for optical signal routing, which is largely analogous to the adoption Wire routing electrical signals. This three-dimensional design allows complex routing mechanisms to operate, completing the necessary steps to mimic the nervous system's operational processes.
The researchers said that the laser is transmitted through the fiber to the chip. Depending on the intensity and distribution pattern of the selected light, the chip routes each input to the output group. To evaluate the output, they produce an image of the output signal. The results show that The final output of this method is highly uniform and the error rate is low, achieving precise power distribution.
The research team said that they really did two things. First, the use of 3D design models to achieve more optical connections in the transmission; In addition, the successful development of new measurement technology enables the characteristics of many devices in the optical subsystem to be reflected. For the large-scale in-depth study of the photoelectron nervous system, these two breakthroughs will play a crucial role.
Editor-in-chief
Neural networks are already popular technologies in the field of artificial intelligence. Whether it is image recognition, face recognition or natural language processing, neural networks are used. Neural networks are indeed powerful, but precisely because of its power, the lowest-level chips are proposed. New challenges. It is foreseeable that traditional chips will one day fail to load the computational requirements put forward by the artificial intelligence era. At this time, the 'light' debut is highly anticipated – using optical circuits instead of circuits, data transmission and processing speed. It has become faster. Many teams are already developing optical chips. However, the results are generally still in the laboratory stage. Towards industrialization, factors such as cost, consistency and stability should be considered.