Intel is one step closer to replicating some parts of the human brain on a computer chip: the company has published new research into an algorithm capable of smelling – or rather, of recognizing what it is smelling – based on the same biological signals that make us jump to the oven when we sniff the familiar smell of burnt pizza.
In partnership with researchers from Cornell University, the technology firm's neuromorphic computing group built a mathematical algorithm that mimics the olfactory systems observed in mammals that are responsible for learning and identifying smells. The algorithm was then implemented on Intel's Loihi neuromorphic computing chip – a 14-nm, 130,000-"neuron", 130 million-"synapse" heavy system, which is based on the design of the brain.
Hence the term "neuromorphic," which refers to the act of trying to make computers think and process information like biological brains. As Mike Davies, director of the neuromorphic computing lab at Intel puts it, neuromorphic computing "picks the actual process of brains and puts that into silicon."
Intel's team started looking at what happens in mammals' brains when they smell – and as it turns out, that's predictably, quite a lot of things. About 450 different types of olfactory receptors sit in our noses, which can be activated by airborne odor molecules to send signals to the brain, where electrical pulses within a connected group of neurons generate the sensation of a particular smell.
What's more, our brains can store past experiences of smells, as well as cross-reference different ones, to constantly process new information and make sure that we distinguish between trillions of different scents.
Intel, of course, is still far from having replicated the brain's entire olfactory system on a chip. Based on the same principles, the company's neuromorphic computing team rather started with building an algorithm that can effectively identify ten different smells.
To do so, researchers recorded the response of 72 chemical sensors sitting in a wind tunnel, as they circulated ten different scents – including methane, ammonia or acetone – in the tunnel. The dataset was then fed to the Loihi chip, which was able to draw neural representations of each of these smells; just like the brain assigns scents to specific patterns of electrical signals.
Nabil Imam, senior research scientist in Intel's neuromorphic computing groups, said: "My friends at Cornell study the biological olfactory system in animals and measure the electrical activity in their brains as they smell odors. On the basis of these circuit diagrams and electrical pulses, we derived a set of algorithms and configured them on silicon, specifically our Loihi test chip."
Loihi is capable of self-learning thanks to a new type of neural network dubbed "spiking neural network". The chip's system, using biologically-realistic models of neurons, can process real-world sensory data to make sense, to a certain degree, of its environment .
As a result, not only was Loihi capable of learning and identifying the scents it was presented with; but it showed that it could smell correctly even in the case of background interference.
Different scents from various sources mingling, said Imam in his research paper, "has been recognized as one of the central problems in neuromorphic olfaction"– but one which Loihi seems to have overcome, at least to a certain extent. Intel effectively reported that the new technology distinguished between smells "even in the presence of strong background interferents."
The chip's capacity to identify odors based on patterns defined by sensors would be a significant step up from existing technologies, like smoke or carbon monoxide detectors, which only react to certain molecules in the air, without categorizing them.
Imam mentioned potential applications for the technology in environmental monitoring: equipping robots with Intel's electronic nose system could enable faster detection of hazardous materials, for instance; or it could help with identifying dangerous substances in airport security lines.
Ultimately, the researcher's goal is to apply neuromorphism to a wider range of use-cases. "My next step is to generalize this approach to a wider range of problems – from sensory scene analysis to abstract problems like planning and decision-making," he said. "Understanding how the brain's neural circuits solve these complex computational problems will provide important clues for designing efficient and robust machine intelligence."
In the quest for a computer mapping of the brain, he is not alone. In Germany, the Jülich Research Centre's Institute of Neuroscience and Medicine is working with the SpiNNaker supercomputer to mimic cortical microcircuits, while IBM's Neuromorphic Devices and Architecture Project is also carrying out similar experiments.
Intel, however, seems particularly dedicated to the advancement of neuromorphic computing. Last year, the company announced that an 8 million-neuron neuromorphic system made of 64 Loihi research chips would be made available to the research community. Code-named "Pohoiki Beach," the system was launched to let scientists scale-up their neural-inspired algorithms.