Look ma, no hands
Chinese chip creates a bicycle that responds to commands and can look after itself. Nick Carne reports.
Chinese scientists have unveiled this autonomous bicycle which, they say, combines brain-inspired and computer-science-based approaches to artificial intelligence.
At its heart is the Tianjic chip they developed at Tsinghua University, Beijing, in 2015.
This hybrid technology, they believe, has the potential to improve the capability of these systems to achieve artificial general intelligence (AGI): a platform that could, in principle, perform any task that a human is capable of.
At this stage, at least, it allows this bicycle to respond to voice commands, keep out of trouble and detect and follow a person.
Luping Shi and colleagues describe their work in a paper published in the journal Nature.
“Using just one chip, we demonstrate the simultaneous processing of versatile algorithms and models in an unmanned bicycle system, realising real-time object detection, tracking, voice control, obstacle avoidance and balance control,” they write.
AI based on neuroscience attempts to closely mimic the brain, while a computer-science approach, uses computers to perform machine-learning algorithms.
The ultimate goal, the researchers say, is to combine the two, but this is made difficult by the fact that the two systems use distinct and incompatible platforms.
“Although both approaches can solve subproblems in specialised domains where data are abundant, it remains difficult to solve complex dynamic problems with the uncertain or incomplete information that is associated with many systems,” they write.
“To further improve the intelligence capability needed to achieve AGI, there is an increasing trend to incorporate more biologically inspired models or algorithms into the prevailing ANNs [artificial neural networks], resulting in a more explicit dialogue between the two approaches.”
Their hybrid chip makes a fair fist of it. It has many functional cores that are highly reconfigurable, allowing various neural networks and hybrid coding schemes to be freely integrated, allowing for “seamless communication among multiple networks”.
Four different neural networks were pretrained and programmed onto the chip, and its decentralised architecture and arbitrary routing topology allowed them to operate smoothly in parallel – which meant the bicycle accomplished all set tasks with relative ease.
“Our research has examined a novel neuromorphic architecture that offers flexibility by integrating cross-paradigm models and algorithms onto a single platform; we hope that our findings will accelerate the development of AGI, with many possible real-world applications,” the researchers write.