In order to smell, our brains and noses have to work together, so the idea of digitizing smell seems pretty “out there.”
However, if you think about it, our noses are sensing molecules. Those molecules can be identified by a computer, and the smells the humans associated with them can be cataloged. It’s not quite teaching a computer to smell on its own, but maybe it’s best we don’t give them too many human abilities.
The Enigma of Olfaction
While we’ve successfully translated light into sight and sound into hearing, decoding the intricate world of smell remains a challenge.
Olfaction, compared to our other senses, is mysterious, diverse, and deeply rooted in both emotion and memory. Knowing this, can we teach machines to interpret this elusive sense?
A collaboration between the Monell Chemical Senses Center and the startup Osmo aimed to bridge the gap between airborne chemicals and our brain’s odor perception. Their objective was not just to understand the science of smell better but to make a machine proficient enough to describe, in human terms, what various chemicals smell like.
Osmo, with roots in Google’s advanced research division, embarked on creating a machine-learning model. The foundation of this model was an industry dataset, which detailed the molecular structures and scent profiles of 5,000 known odorants.
The idea? Feed the model a molecule’s shape and get a descriptive prediction of its smell.
That might sound simple, but the team had to make sure they could ensure the model’s accuracy.
The Litmus Test: Man vs. Machine
To validate the machine’s “sense of smell,” a unique test was devised.
A group of 15 panelists, trained rigorously using specialized odor kits, was tasked with describing 400 unique odors. The model then predicted descriptions for the same set.
Astonishingly, the machine’s predictions often matched or even outperformed individual human assessments, showcasing its unprecedented accuracy.
Machines That Can ‘Smell’ vs. Digitizing Smell
Beyond its core training, the model displayed unexpected capabilities. It accurately predicted odor strength, a feature it wasn’t explicitly trained for, and identified distinct molecules with surprisingly similar scents. This accomplishment suggests we’re inching closer to a world where machines can reliably “smell.”
But for now, that’s overstating it. The team has made a major leap towards digitizing smell. But machines don’t have senses. They can only replicate the kind of information our brains produce when we smell things. Of course, they don’t have any sense of enjoyment (or repulsion) at certain smells.
In any case, the Monell and Osmo collaboration has significantly advanced our journey in understanding and replicating the sense of smell. As we move forward, this research could revolutionize industries from perfumery to food and beyond.