Advancing AI’s Sense of Smell

Smell is a mystery to scientists, unlike our ability to understand sight and hearing. We can describe what we see in terms of light wavelengths and what we hear in terms of sound waves, but smell remains uncharted territory. In recent research featured in the journal Science, scientists used a neural network to teach a computer about 5,000 compounds from perfume databases.

These compounds are molecules that have distinct smells, like “fruity” or “cheesy.” The AI created a “principal odor map” that visually represented how different smells relate to each other. When the scientists introduced a new molecule to the AI, it could predict what it would smell like in descriptive terms.

The research team went a step further by having 15 adults from diverse racial backgrounds in the Philadelphia area smell and describe the same odor. What they discovered was that, most of the time, the AI’s descriptions were more accurate than those of the average panelist.

Alex Wiltschko, one of the authors of the study and the CEO and co-founder of Osmo, a company dedicated to giving computers a sense of smell, emphasized this achievement. Osmo collaborated with researchers from Google and several US universities on this groundbreaking research.

The study’s breakthrough in teaching AI to understand and describe smells is a significant advancement in our understanding of scent perception. While scent is highly personal and influenced by individual memories and culture, this AI-driven approach has managed to capture common aspects of smell perception rooted in chemistry.

The ability of the AI to produce an odor map and predict the scents of new molecules is a noteworthy achievement, even though it is not the first of its kind. This research opens up exciting possibilities for various applications, from fragrance design to potential advancements in fields like healthcare and environmental monitoring.