A COMPUTER chip inspired by the brain could pave the way for artificial intelligence with a broader range of abilities. The chip has been used to control an autonomous bicycle, but one day it might power self-driving cars and smart robots.
Shi Luping at Tsinghua University in China and his colleagues made the chip with hardware based on the structure of the human brain that can run two types of algorithm.
Advertisement
To date, most approaches to developing AIs fall into two camps. The first, and more common, type of algorithm is artificial neural networks, which are simplistic models of neurons designed for specific computer processing tasks such as recognising objects.
鈥淥ne day, this flexible computer chip might power self-driving cars and smart robots鈥�
Then there are biologically inspired circuits, known as spiking neural networks, which are rarer. These aim to more closely replicate the human brain, with its billions of neurons that communicate via electric pulses. Whereas neurons in artificial neural networks can 鈥渇ire鈥� at different intensities, neurons in spiking neural networks are either on or off.
Shi鈥檚 chip is built to deal with the operations required by both algorithms, so it is potentially more efficient.
His team has shown how the hybrid chip could be used in an autonomous bicycle. The bike combines several different algorithms for different tasks, allowing it to respond to voice commands, to self-balance, avoid obstacles, and detect and follow a person.
[video_player id=鈥滾QVDOqtE鈥� access_level=鈥漞veryone鈥漖
Its processor, known as the Tianjic chip, combined five different neural networks, including a spiking neural network that recognises voice commands such as 鈥渟traight鈥� and 鈥渟peed up鈥� (Nature, ).
The algorithms were all trained before being loaded onto the chip, where they were run in parallel in the bicycle. Shi says the computer chip could be used for other applications including intelligent robots and self-driving cars.
Alessandro Oltramari at engineering firm Bosch says the hybrid chip may be suitable for a variety of tasks. 鈥淵ou really need to have an architecture that is complex enough, that can deal with different problems using different mechanisms,鈥� he says.
But more general AIs are still some way off, he adds. Although the Tianjic chip was effective for autonomous navigation in a bicycle, to prove its general ability, it would need to succeed at a range of totally different tasks, such as manipulating objects, playing games or holding a conversation.
Oltramari says the chip overlooks a third approach that may be required for artificial general intelligence: cognitive-inspired neural architectures. Instead of trying to replicate the anatomical structures of the brain, this aims to mimic the cognitive abilities of the human mind, such as its ability to pay attention, remember and make predictions.
