From recognising friends in your photos to plastering an Instagram filter over your face in real-time, you don’t have to look far to see that computers are now capable of performing complex image processing. These impressive feats are possible largely due to the development of artificial neural networks, a circuit of components that mimic the behaviour of neurons and synapses in the brain. Just like their biological counterparts, however, artificial neural networks do have their shortcomings—they consume a lot of energy.
Memristor, a type of electrical component, might just be the solution to less energy-intensive artificial synapses for neural networks. Memristors arranged in a grid, forming what is known as a crossbar array (CBA), hold promise for building scalable neural networks with high-performance computing capacities.
This is particularly true for memristors made from two-dimensional (2D) materials which possess unique properties and superior device performance compared to their transition metal oxide-based counterparts. But due to limitations in fabrication methods, 2D material-based memristors aren’t as easy to make and integrate into devices.
Now, a team of researchers led by Dongzhi Chi and Kah-Wee Ang, Principal Scientists at A*STAR’s Institute of Materials Research and Engineering (IMRE), has demonstrated a novel way to create and integrate a 2D memristor CBA into neural network hardware—without the drawbacks.
“Previous research focussed mainly on single device characteristics, which is far from array-level applications,” said Chi and Ang. “Here, we fabricated memristor CBAs and demonstrated its capacities at an array level.”
The researchers used a process called molecular beam epitaxy to create wafer-scale 2D hafnium diselenide (HfSe2) films, the material for their memristors. Unlike the inconsistent mechanical exfoliation process commonly used to fabricate 2D materials, the researchers were able to achieve controllable and uniform growth of the films.
Furthermore, the use of ultra-thin HfSe2 films allowed filaments to form within the memristors in the CBA at a low voltage, a key feature for facilitating signal processing with low switching energy.
“We leveraged the intrinsic defects that exist in the polycrystalline HfSe2 film to facilitate the formation of filament for switching,” explained the researchers. “The result was a memristive CBA with a low switching voltage and thus lower energy consumption.”
When the researchers tested their creation on the multiply-and-accumulate (MAC) operations widely used in image processing, they found it performed with a high level of efficiency and accuracy.
“The CBA can also perform multiple MAC operations simultaneously, thereby enabling parallel computing,” added Chi and Ang.
The team hopes to integrate their CBAs with access devices such as selector diodes that will allow precise control of individual memristors. They also plan to build a complete image processing hardware by connecting their CBAs with custom-designed circuits.
The A*STAR-affiliated researchers contributing to this research are from the Institute of Materials Research and Engineering (IMRE).
