Geoffrey Hinton outlines new AI for pattern matching that doesn't eat up data
Hinton, working with Google, claims new approach to AI could be applied to speech recognition
Geoffrey Hinton, the cognitive psychologist and computer scientist, most noted for his work on neural networks, has unveiled a new AI advancement that, he claims, won't over-consume large datasets.
Hinton, who now works for both Google and the University of Toronto, has found a way for the technology to correctly identify images while relying on less data.
Specialising mainly in artificial neural networks and pioneering the commercialisation of machine learning, he's published his findings in two research papers.
Computers used for image recognition usually have a database of existing images, but Hinton's approach could result in AI being able to identify faces from different angles.
In addition, the technology, he suggested, could also be applied to speech and video recognition. Addressing the Go North Technology Summit, he said: "This is a much more robust way of identifying objects, according to Reuters.
The summit was hosted by Google, and he spent much of his talk proving the theory he originally devised in 1979. He addressed the work he conducted with Google researchers Sara Sabour and Nicholas Frost.
During their research project, the specialists deployed individual capsules, or groups of virtual neurons, to explore their relationship. Artificial neural networks attempt to mimick the way the human brain works.
The researchers wanted to see if the system they devised would be able to confirm whether specific features were presented by images completely new to the system. The results were not only positive, but trumped existing image recognition technologies, Hinton claimed.
Hugo Larochelle, who heads up Google Brain's research efforts in Montreal, Canada told Reuters: "The hope is that maybe we might require less data to learn good classifiers of objects, because they have this ability of generalising to unseen perspectives or configurations of images,"
"That's a big problem right now that machine learning and deep learning needs to address, these methods right now require a lot of data to work."