Fingernail-sized chips or replace supercomputers?

According to the official website of the Ministry of Science and Technology, a few days ago, MIT engineers designed an artificial synapse that can precisely control the intensity of current flowing through it, similar to the way that ions flow between neurons, and has used silicon germanium. Made of artificial synapse chips. The chip and its synapses can be used to identify handwritten samples in simulation studies with an accuracy rate of up to 95%. The research results are published in the journal Nature-Materials, which indicates that humans are portable and low-powered. The consumption of neuromorphic chips has taken an important step.

It is reported that there are about 100 billion neurons in the human brain. Neurons pass instructions through 100 trillion synapses, enabling the brain to identify patterns with lightning speed, complete memories and perform other learning tasks. Emerging areas 'neurological form calculation' The researchers tried to design a computer chip that works like a human brain and works by simulating signals similar to neurons. In this way, small neuromorphic chips can handle millions of parallel computations as efficiently as the brain. At present, only large-scale supercomputers can be realized. The urgent problem to be solved in this portable artificial intelligence method is synapse.

Researchers use silicon germanium to make neuromorphic chips composed of artificial synapses. Each chip is composed of 'input/hide/output neurons'. Each neuron is connected to other 'neurons' through filament-based artificial synapses. Each synapse is about 25 nanometers, and the difference in ion current between them is only 4%. It is the most consistent device that can be achieved in the laboratory, and it is also the key to demonstrating artificial neural networks. The researchers then conducted artificial neural network computers. Simulation, recognition of handwritten samples, the accuracy rate reached 95%, while the accuracy of the existing software algorithm is 97%.

The team is developing simulation-based neuromorphic chips that can be used to identify handwriting tasks, and expects to use its artificial synapse design to create smaller, portable neural network devices for performing complex calculations, eventually realizing the use of fingernail-size chips instead supercomputer.

2016 GoodChinaBrand | ICP: 12011751 | China Exports