Communications of the ACM
What's That Smell?
The new system, which runs artificial intelligence software on Intel's Loihi neuromorphic chip, is essentially an "electronic nose" that can learn the scent of a chemical from a single exposure.
Credit: Rambus
Researchers at Cornell University and Intel have developed artificial intelligence (AI) software that can learn the scent of a chemical with just one exposure, and then remember that scent forever.
The software, which is designed to run most efficiently on an experimental chip from Intel known as Loihi, is so precise, it can even detect a scent that's masked by a number of other scents, according to researchers.
Ultimately, the researchers hope to produce a market-ready solution that can detect hazardous substances in the air, sniff out dangerous drugs, discover hidden explosives, and assist with medical diagnoses.
"Low-energy modules built around Loihi, running our algorithm, and hooked-up to diverse sensor arrays could be built into robots, medical analysis devices; for example, blood composition, hyperspectral processors, air quality sensors, food processing pipelines, you name it," says Thomas A. Cleland, a member of the research team and associate chair and professor of psychology at Cornell University.
The system works by processing an input signal pattern for a scent drawn from an array of sensors, then recording that signal pattern in the AI software as a recognizable scent for future use.
The beauty of the system is that a scent detected by the software only needs to be 'learned' once, and the system will always be able to identify that pattern, according to Cleland.
Another advantage of the software is that all learning regarding a scent is 'local.' Essentially, when the software discerns a scent, only a few cyber-synapses (the AI equivalent of human synapses) are encoded. "All learning is local to particular synapses and derived only from things that happen at that synapse," Cleland says.
That approach differs markedly from conventional deep learning, which would require the software be exposed to hundreds or even thousands of samples of the same scent before it could discern a common pattern it could use as a reference for that scent.
The system's AI also differs from traditional deep learning in that the software can add new scents to its repertoire without losing the memory of a previous scent. This technique also offers a marked advantage over conventional deep learning systems, which generally need to 'forget' a previously learned scent when prompted to learn a new scent, according to Cleland.
Under the hood, the system relies on data generated by 72 chemosensors, which send a fingerprint signal for each of 10 scents (including ammonia and carbon) to the system's software. Once detected by the AI, the fingerprint signal is processed, recorded, and stored, enabling the system to instantly recognize that same scent the next time it is sensed.
Processing is greatly aided by Intel's experimental Loihi neuromorphic chip system, a network of 64 chips offering the power of 8 million neurons, which Cleland describes as "ridiculously energy-conservative."
Tests showed that even when a sought-after scent was overpowered by a number of other scents—what Cleland calls 'noise'—the software was still able to ferret-out the scent sought.
"This classification under noise is really our killer app," Cleland says. "The visual analogy is that you can't see something if there is an obstruction in the way, but maybe you can recognize your friend just from a corner of his head and one ear poking out from around the side of this visual obstruction, because you know what your friend's head looks like, even if most of his head is not able to be sensed directly because of that obstruction."
Cleland says development of the AI software was initiated at his lab, with the help of Nabil Imam (while Imam was a Ph.D. student at Cornell), along with help from current Cornell Ph.D. student Ayon Borthakur. "Nabil got hired by Intel a few years after his Ph.D. defense, after some time at IBM, and joined the Intel team that developed that new Loihi chip," Cleland recalls, "so we ported the algorithm to the Loihi chip for its debut publication."
It was a win-win collaboration, according to Cleland: "Extra buzz for the algorithm, a cool application for the Loihi chip; everybody is happy."
Even so, the system needs additional work before it can be commercialized, including studying how the AI software performs 'in the wild' under completely unpredictable conditions.
Says Santiago Marco, group leader of the Artificial Olfaction Lab at Spain's Institute for Bioengineering of Catalonia, who specializes in signal and data processing for sensing systems, "I would like to see how this research can be extended to do the odor classification tasks in real time and with a larger diversity of conditions."
Timothée Masquelier, a researcher specializing in bio-inspired AI at the French National Centre for Scientific Research, agrees on the need for further research. "The proof of concept is convincing, but the problem they tackle is somewhat simple: identification of a previously learned odor, despite noise.
"Most real-world problems will involve recognition of broader odor categories; for example, flowers, food, animals, etc. It remains to be proven if their system can handle this sort of categorization."
Adds Brice Bathellier, an experimental and theoretical neuroscientist at the French National Centre for Scientific Research, "It is not ready to be a real flexible nose, simply because it has been tested with too few odors, in very controlled conditions. But it seems to work to identify some odors among variable mixtures. Overall, I think this is a nice algorithm for the problem of odor recognition with this type of chemical chip."
Enhancing scent detection and performing experiments in the wild will hinge on linking the system to a greater number of sensors, Cleland says. "We really need collaborators with larger sensor arrays to hook up to our" network to move forward with research.
Fortunately, a wide variety of sensors will work for further development. "A great advantage of this algorithm, as inspired by the history of work in the artificial olfaction community, is that it doesn't need very specialized sensors to be developed to detect this or that particular signal or compound," Cleland says. The system works "on any sensor array that can be processed as a list, not just a chemosensor array."
"On the hardware side, we're looking to Intel," Cleland says. "We can run this on the Pi (computer). But the Loihi chip is intended for exactly this sort of purpose."
Meanwhile, Imam, now a senior research scientist at Intel, also is looking to repurpose the research other types of analysis. "My next step is to generalize this approach to a wider range of problems, from sensory scene analysis—understanding the relationships between objects you observe—to abstract problems like planning and decision-making.
"Understanding how the brain's neural circuits solve these complex computational problems will provide important clues for designing efficient and robust machine intelligence."
Joe Dysart is an Internet speaker and business consultant based in Manhattan, NY, USA.
No entries found