Artificial intelligence is being incorporated into nearly everything but running such systems creates an insatiable demand for energy. This is why tech companies and researchers across the globe are actively trying to develop neural networks that consume less power.
To this end, IBM has developed a new silicon-based chip which contains all the key features of a neural network and is a 100 times more energy efficient than the existing chips at present.
Neural networks, which is inspired by how the human brain works, is usually written in software and integrated with a device. Creating neural nets in software is quite easy but reproducing them with hardware has been difficult to achieve, until now.
IBM researchers have come up with microelectric artificial synapses on chips that can mimic the synapses that connect individual neurons in the brain. Published in the journal Nature, the paper describes neuroscience as the inspiration for this project where they’ve two types of synapses – short-term ones for computation and long-term ones for memory.
This method addresses several critical issues including low accuracy which has foiled previous attempts to build artificial neural networks in silicon. The researchers tested the neural networking chip by assigning two simple image recognition tasks related to handwriting and color image classification.
The results were astonishing as they found the system to be as accurate as a software-based neural net and the fact that it consumed only 1 percent energy in comparison.
“A factor of 100 in energy efficiency and in training speed for fully connected layers certainly seems worth further effort,” said Michael Schneider, a researcher at National Institute of Standards and Technology.
However, IBM has a long way to go as the design of its chip is quite bulky. It carries five transistors and three other components in comparison to a single transistor on existing standard chips. The company also needs to perform more tests as certain aspects of the system have been only simulated for testing.
Nevertheless, this development has its own significance as it could pave the way to advanced computers with artificial intelligence logic embedded into its core. Such an achievement will be crucial not only for the AI domain but also for hardware sector where companies like IBM are focusing on reinventing computer hardware.