Visual artists can consult a digital color wheel to test relationships between different shades. Audio engineers can alter soundwave frequencies to create complex tones and harmonies. But no digital tools exist that can predict how people will perceive the blend of scents wafting from a steaming cup of coffee or a freshly baked apple pie.
In new Yale-led study, researchers developed a machine learning method to accurately predict how people perceive different scent combinations. The advance is an important step toward creating a quantitative framework for understanding our perceptions of complex odors, which the researchers say could someday lead to applications relevant to human health.
Is it possible to measure how people perceive odors?
Unlike vision, where the trichromatic theory helps explain how combinations of primary colors enrich our visual world, there is no quantifiable framework for mapping our perceptions of odors. A new Yale-led study uses machine learning to predict how people will perceive complex odors — an important step toward digitally mapping scent perceptions.
What are the potential benefits of digitally mapping perceptions of smell?
Potential applications could be relevant to human health. Many health conditions, such as Parkinson’s and cancer, have been reported to have odor signatures. There could be technology years from now that can monitor your smell and alert you to any changes that might require additional screening.
“Much progress was made over the past decade in identifying how individual molecules affect perceptions of odor, but the smells we encounter in the real world are composed of complex mixtures of dozens or hundreds of molecules, and we lack a metric for comparing and measuring them,” said lead author Vahid Satarifard, a research scientist at Yale’s Human Nature Lab. “This study shows that perceptual distances between odor mixtures can be accurately mapped, laying the groundwork for digitizing sense of smell.”
Thestudy was published in the journal Proceedings of the National Academy of Sciences. Nicholas Christakis, Sterling Professor of Sociology and Natural Science in Yale’s Faculty of Arts and Sciences and director of the Human Nature Lab, is a coauthor of the study.
The research was part of a Dialogue for Reverse Engineering Assessment and Method (DREAM) olfaction prediction challenge — an international competition organized by Pablo Meyer from IBM research — in which teams were invited to develop predictive models to measure perceptual similarities in scents produced by pairs of molecule mixtures. The contest builds on a previous DREAM challenge that had demonstrated that computer models could predict perceptual descriptors like “garlic,” “sweet,” and “floral” for individual molecules.
For the new challenge, the organizers gathered data from three prior studies on odor-similarity measurements into a unified dataset, comprising 168 unique single molecules, 731 unique molecule mixtures, and 507 mixture-pair measurements. Over roughly three months, 26 teams competed to minimize prediction errors on a hidden test-set of 46 scent mixture pairs.
Their models were meant to accurately predict how similar two pairs of odor mixtures will smell to humans, Satarifard explained.
Four teams tied for first place, including the team from the Human Nature Lab, which subsequently led the post-challenge phase of the research project.
After the competition, the researchers created an ensemble model by averaging the predictions of the four top-performing models along with those of two other high-performing models from the challenge.
“The ensemble model is essentially a machine-learning instance of the ‘wisdom of crowds’ phenomenon,” Satarifard said. “It’s the idea that the collective intelligence of a group — in this case, these six high-performing models — is typically superior to that of any single model.”
The ensemble model outperformed existing state-of-the-art approaches to predicting complex scent perceptions, as well as the top-performing competition models, on the hidden test set. It also maintained a strong performance against an independent dataset of 50 newly designed scent-mixture pairs, the researchers said.
The highest performing models relied heavily on semantic descriptions of odors rather than on their molecular structures, the study found.
“We found that using language features was very powerful in predicting similarity between two scent mixtures,” Satarifard said. “This is interesting because English has a small vocabulary for smell, and we usually describe odors by naming objects. Something ‘smells like flowers’ or like watermelon or like a rotten egg, whereas colors have specific names like ‘green’ or ‘blue.’ That was thought to make semantic odor descriptors a poor basis for predicting how smells relate, but we found the opposite.”
The finding contributes to an ongoing discussion about using language, rather than chemical composition, to describe and model smell, Satarifard said.
The research helps to open the door to potential practical applications, including digital olfaction and mapping scents that are relevant to people’s health, Satarifard said.
“Many health conditions, such as Parkinson’s and cancer, have been reported to have odor signatures,” he said. “One ultimate goal here is to reach a place where we can use odor as a disease biomarker. There could be technology several years from now that can monitor your smell and alert you to any changes that might require additional screening.”
A further use of this technology is to assess the odors emitted by people in non-clinical ways, Christakis noted.
“Body scent is also a complex mixture of odors,” he said, “and we suspect that it plays an important role in human social interactions.”
