A new type of computer-trained odor model does a better job of identifying odors than humans. When analyzing 500,000 potential odor molecules that have never been synthesized, it also easily does what would take 70 man-years to complete. While machines are getting better at mimicking human vision and taste, they're lagging behind in developing their sense of smell.
Of course, there are already electronic noses that can sniff cancer cells in bleeding cells and assess the air around wastewater treatment plants, but a truly computer-driven sense of smell has been elusive. This may be because our nose has 400 olfactory receptors, which is a lot more than the 4 receptors of vision and about 40 receptors of taste.
Researchers at the University of Pennsylvania's Monell Center for Chemical Senses, along with colleagues at Osmo, a spin-off of Google DeepMind, led a study to create a neural network-based system that analyzes odor molecules and describes in human language what that molecule should taste like. The AI system developed what the researchers call a "master odor map" (POM).
"In olfactory research [...] it has been a mystery what physical properties allow airborne molecules to produce such odors in the brain." But if computers can discern the relationship between the shape of a molecule and how we ultimately perceive its smell, scientists can use this knowledge to deepen our understanding of how our brains and noses work together.
This knowledge could help researchers develop better mosquito repellents or deodorant products, among other possible applications.
To train the system, the team fed it the molecular structure of 5,000 odorous substances, as well as a series of descriptions describing odors, such as "mint" or "musty." The team also brought in 15 panelists, asked them to sniff out 400 scents, and gave them 55 words to describe each smell.
In the tests, the AI system performed slightly better than the panelists. But there is also a more impressive result.
"The most surprising result, however, was that the model successfully performed the olfactory task for which it was not trained," Mainland said. "It was eye-opening that we never trained it to learn odor intensity, but it was still able to make accurate predictions."
Next, the researchers used the system to map 500,000 odor molecules that had never actually been synthesized — a task that the team says would take humans 70 years to sniff.
"Advances in neuroscience are often measured by the creation and discovery of neural circuit-enabled maps of the new world," the researchers wrote. This was possible because the scientists first had a map of the outside world and then measured how responses in the brain changed depending on where the stimulus was on the map. This study proposes and validates a data-driven human olfactory map. We hope this map will be helpful to researchers in chemistry, olfactory neuroscience, and psychophysics [...] Become a new tool for studying the nature of olfaction. "
The study was published in the journal Science.