Chinese researchers from Fudan University have created a device, based on an ultrathin semiconductor monolayer molybdenum disulfide (MoS₂), capable of simulate real neurons in the brain.
With the development of AI systems and machine learning, there is a growing demand for hardware components to speed up data analysis and reduce power consumption. Since machine learning algorithms are inspired by the work of neural networks in the human brain, a number of engineers are creating hardware, that will mimic the work of neurons in the brain.
A neuromorphic hardware architecture typically includes interconnected artificial neurons. Over time, the connection between them can weaken or, on the contrary, become stronger. This process resembles synaptic plasticity, the ability of the brain to adapt over time in response to experience and learning. By mimicking synaptic plasticity, neuromorphic computing systems are able to execute machine learning algorithms more efficiently, consuming less power and analyzing larger amounts of data.
The device, created by chinese scientists, combines dynamic memory with random access memory (DRAM) and metal-on-silicon (MoS₂) chips. According to the leading authors of the study Yin Wang, Saifei Gou and their colleagues, neuromorphic equipment that accurately mimics. The diverse behavior of neurons can be useful for the development of advanced artificial intelligence systems.
“Hardware that incorporates synaptic plasticity — adaptive changes that strengthen or weaken synaptic connections — has already been investigated, but mimicking the full range of learning and memory processes requires the interaction of multiple plasticity mechanisms, including intrinsic plasticity. We show, that a neuron operating on the principle of integration and activation can be created by combining dynamic random-access memory and a wafer-scale inverter based on a monolayer of disulfide films”, — the researchers explain.
The developed artificial neuron consists of a dynamic random access memory (DRAM) on an inverter circuit. DRAM is a memory system capable of storing electrical charges in capacitors. The amount of electric charge in capacitors can be modulated to simulate changes in the electric charge on the membrane of biological neurons, which ultimately determines their activation.
An inverter is an electronic circuit, capable of converting a high voltage input signal into a low voltage signal and vice versa. In an artificial neuron, this allows you to generate electrical impulses similar to those, that occur in real neurons when they are activated.
“In this system, the voltage in a random-access dynamic memory capacitor, i.e., the potential of the neuron’s membrane, can be modulated to simulate internal plasticity. The module can also mimic the photopic and scotopic adaptation of the human visual system, dynamically adjusting its light sensitivity”, — the authors of the study emphasize.
To assess the potential of the created artificial neuron, the researchers created several samples and assembled them into a 3×3 array. After that, they tested the ability of this neural network to adapt its own responses to input signals depending on changes in lighting, imitating the adaptation of the human visual system to different lighting conditions. The artificial neuron has already demonstrated its effectiveness in computer vision and image recognition models, while consuming significantly less power.
“We created a 3×3 array of photoreceptor neurons and demonstrated light coding and visual adaptation. We also use the neural module to model a bio-inspired neural network model for image recognition”, — the researchers emphasize.
The results of the study are published in the journal Nature Electronics
Source: TechXplore
Spelling error report
The following text will be sent to our editors: