A new computer chip inspired by the human brain could fix some of the biggest problems with artificial intelligence.
The system – built around a “memristor” that mimics the way that neurons are connected in the brain – could dramatically reduce the amount of energy used by AI, as well as helping it to learn in the same way we do.
The nanoelectronic device is based on a newly developed kind of hafnium oxide that is able to work as a stable and low energy component, mimicking the human brain.
Current AI systems use vast numbers of conventional computer chips. They have to shuttle data back and forth between their memory and processing units, which not only leads to the vast amount of energy used by AI but also limits their functioning.
Researchers believe that new systems based on the form of the human brain could allow them to dramatically reduce that energy use, by as much as 70 per cent. They do so by storing and processing information in the same place, using extremely low power as well as being adaptable in the way our brains are.
“Energy consumption is one of the key challenges in current AI hardware,” said lead author Babak Bakhit, from the University of Cambridge, in a statement. “To address that, you need devices with extremely low currents, excellent stability, outstanding uniformity across switching cycles and devices, and the ability to switch between many distinct states.”
So far, work on building these human-inspired chips has relied on tiny conductive filaments that are inside a metal oxide. But they unpredictable, and require large amounts of electricity, meaning that it is hard to make them useful.
The new system, developed at Cambridge, uses the new kind of material to make a film that is able to switch states differently. Researchers were able to make tiny electronic gates inside oxide, allowing the device to change resistant smoothly rather than in the more dramatic and power-intensive way currently used.
Tests show that the system could endure tens of thousands of these switches every day, and store their programmed state. They also seemed to adapt in the same way that biological systems do.
“These are the properties you need if you want hardware that can learn and adapt, rather than just store bits,” said Dr Bakhit.
There are still some problems with the system, such as the fact that making them requires huge amounts of heat. But the researchers hope to bring that temperature down and put the devices onto a chip.
The work is reported in a new paper, ‘HfO2-based Memristive Synapses with Asymmetrically Extended p-n Heterointerfaces for Highly Energy-efficient Neuromorphic Hardware’, published in the journal Science Advances.
