|Google's Research Team taught an AI to smell using a set of about 5,000 molecules from perfumers / Photo Credit: Ociacia (via Shutterstock)|
Google may have its own perfume, but it’s not going to market it anytime soon because it wants to tap into a human’s sense of smell, wrote Sara Harrison of monthly American magazine Wired. Researchers at Google Brain published a paper on preprint site Arxiv presenting how they trained an AI to discern smells.
Interestingly, the science of smell is behind other fields such as light, which has been understood for centuries. Hence, Google contributed a drop of knowledge to the existing pool of knowledge on the science of smell. Molecules are the foundation of scents. When a scent enters a person’s nose, the molecules will interact with receptors. Then, the receptors will send a signal to the olfactory bulb, a small part in the individual’s brain, prompting them to identify the smell. If the scent smells like popcorn, they will say, “Mmm, popcorn.”
It is difficult to identify a molecule’s odor from its chemical structure. Changing or removing one atom or bond will result in odors drastically altering from roses to rotten eggs, according to Alex Wiltschko, who led Google’s research team for the project. Wiltschko’s team used a graph neural network (GNN) for their research.
GNN looks at graphs, akin to how social media websites predict who you should befriend next on the platform. Likewise, GNN processes each of the molecule’s structure “and understand that in one molecule.” To illustrate, the network could understand that a “carbon atom was five atoms away from a nitrogen atom.”
Google’s research team used a set of about 5,000 molecules from perfumers who have an excellent sense of smell to describe each molecule. The team used two-thirds of the data to train the network and tested it to see if it could predict smells. And it did. However, a researcher at Cold Spring Harbor Laboratory pointed out that it’s unclear whether or not scientists can learn something about human olfaction from AI considering the network differs from that of a human’s olfaction system.
Moreover, the research doesn’t say much about “mixtures of combinations of scents.” But if researchers found out what properties and patterns cause molecules to smell a certain way would be “quite an incredible feat,” smell researcher at the Monell Chemical Senses Center Johannes Reisart said.