The sense of smell is an essential part of life. From whiffing freshly washed bedsheets to sniffing burning food in the oven, humankind depends on its sense of smell for both pleasure and practical purposes. Just as artificial intelligence has learned to speak and move, scientists are working on teaching AI how to smell.
Google scientists have made significant progress in this challenging endeavor. Understanding smell is complicated because of how intricate the process is. For reference, when the eye determines a color, it only has sensory receptors for the three primary colors – red, green, and blue. Yet the nose has more than 300 hundred receptors for odors. Making things more complicated, each scent comprises several hundred volatile molecules. Due to the sheer volume, no universal odor map has been created. Another difficulty with mapping odors is that smells are subjective.
However, scientists at Google are not balking at this challenge. In 2019, they developed a graph neural network model that “began to explore thousands of examples of distinct molecules paired with the smell labels that they evoke.” They explain on their blog that this was “to learn the relationship between a molecule’s structure and the probability that such a molecule would have each smell label.” They recently released information on the Principal Odor Map they created. The POM “has the properties of a sensory map.” The POM has been successful thus far. When tested, it excelled with each task. It performed exceptionally well when asked to predict odor, determine the strength of a smell, and what it would smell like to animals.
They are hoping that this odor map can help detect diseases, help to solve problems with food and fragrances, and assist in monitoring the environment. As technologies continue to advance and scientists make breakthroughs, it is only a matter of time until AI will be able to stop and smell the roses.