Artificial Neural Networks made from memristors for brain-inspired computing

January 12, 2023

In a new study by the University of Groningen’s CogniGron Center, scientists explored the potential of a new material designed for neuromorphic computing use. This innovative material features a network of electrically conducting nanoscale channels known as memristors, similar to the brain’s neurons.

Their research paper “Ferroelastic Domain Walls in BiFeO3 as Memristive Networks” was published in the “Advanced Intelligent Systems” scientific journal on November 10, 2022.

*Out of plane (OOP) conduction map of memristors*

What are memristors?

Memristors are a type of passive two-terminal electronic component that can change its resistance based on the amount of electrical charge that flows through it. They combine the functions of memory (“mem-”) and resistors (“-ristor”).

The resistance of a memristor can be programmed and can be used to represent binary states, similar to a traditional electronic memory cell (e.g. DRAM). As they may combine data processing and storage utilizing the same device, they are of interest for research of next-generation computers.

Artificial neural networks based on memristors

Memristors have the potential to be used in a wide range of applications. They have several advantages over traditional memory technologies, including lower power consumption and faster switching speeds. They have been researched as a possible replacement for flash memory, and also for use in other applications such as non-volatile memory and neuromorphic computing.

Neuromorphic computing is a type of computing that is inspired by the structure and function of the human brain.

The human brain and the Artificial Neural Networks (ANNs) are both based on the concept of interconnected nodes or neurons, that process and transmit information. However, there are several key differences between the two.

*In-depth comparison of human nervous system and artificial neural system in neuromorphic devices*

Despite the ANN, inside the human brain the neuron junctions called synapses can strengthen and weaken depending on the environment.

In comparison to the ANN, the strength or degree of synaptic connection between neurons is often maintained throughout time, even in the absence of new information. This gives us the ability to learn over time.

The brain works as an in-memory processor where data is processed and stored in the same biological tissue. In conventional computers, the memory and the processor are two separate units, with data being transferred back and forth, slowing down the performance.

Since the memristors have a memory, they may be able to simulate the biological behavior of synapses, allowing ANNs to retain their learned state even when power is removed.

Research method and result

The research team used a ferroelectric material, which is made of countless small dipoles-simple pairs of positive and negative electrical charges separated by very small distances with the size of a few atoms.

During its production, the material forms microscopically small regions in which all dipoles are aligned in the same direction. These regions (called domains) are separated by domain walls (DW). Compared to the rest of the material, DWs have substantially better electrical conductivity.

Finally, the researchers created a network of nanoscale DWs with memristive properties suitable for memory and neuromorphic applications.

Conclusions

This is a new step towards the creation of brain-inspired computer architecture. The results of the researchers’ work offer an insight into using memristors as a promising technology for building artificial neural networks (ANNs) that are designed to mimic the behavior of biological neurons and synapses.

Memristors may enable the development of more efficient and powerful computing systems, but they also have limitations and challenges that need to be overcome in order to fully realize their potential.

Some of the main challenges of the memristors include their limited lifetime, difficulties of fabrication, the complexity, reliability and difficulty to integrate them with the traditional CMOS technology.