Enhancing neuromorphic computing using memristors

Neuromorphic computing is based on the neural network architecture of our brains. But conventional microelectronic technology in standard computer hardware has not been able to replicate the brain’s high energy efficiency and scaling potential.

The brain is extraordinarily energy efficient, because information exchange and processing are event driven, with spiking energy consumed only when and where it is needed. In addition, neurons and synapses are grouped within highly interconnected neural networks, where each neuron is connected on average to 10,000 other neurons.

According to Ielmini et al., resistive switching random access memory devices, also known as memristors, offer a range of physical phenomena that could contribute to new circuit systems that better mimic the brain, with low power consumption within a compact, scalable architecture.

Memristors change their resistance state via permanent modification of the active material, serving as scalable nonvolatile memories for standalone and embedded memory devices. Their electrical transport, switching, and ion migration properties can be used to approximate neuromorphic functions, such as neuronal integration, fire, oscillations, dendritic filtering and synaptic plasticity according to various spike-time, spike-rate learning rules that have been observed in the brain.

Many of these brain-like phenomena in memristors have been individually demonstrated experimentally. Investigating their combination in full neural networks and their extension to alternative architectures, such as multiterminal devices and bottom-up nanostructures, may further develop neuromorphic devices into a mature technology for manufacturable cognitive computing hardware.

“Memristors have attracted strong research interest partly due to the simple structure that allows for a relatively straightforward fabrication and easy integration and scalability within conventional frameworks,” author Daniele Ielmini said. “Moreover, they can be built into 3D-array stacking to enhance high synaptic density.”

Source: “Brain-inspired computing via memory device physics,” by D. Ielmini, Z. Wang, and Y. Liu, APL Materials (2021). The article can be accessed at https://doi.org/10.1063/5.0047641 .