Scientists make progress toward digitizing smell using AI
Over 400 olfactory receptors in your nose translate the estimated 40 billion odorous molecules in the environment into an even greater number of different odors that your brain can recognize. But you hardly ever learn how to describe smells. Most of us are unable to communicate with our sense of smell, in part because we have disregarded it.
According to this article, these limitations are not unique to humans. We have created devices that can “see” and “hear”. Computers express colors using three numbers, the red, green, and blue (RGB) values, which correspond to the different kinds of color-receiving cells in our eyes. A pitch, which determines the pitch of a musical note, is a single number. An image is a map of pixels, while a song is a series of sounds. Yet, a device that is perfect for odor detection, odor storage, and odor reproduction has never been created.
To remedy this, scientists are making efforts. Researchers presented a model that can explain the scent of a molecule as well as, or even better than, a person in a report that was released at the end of August (at least in limited trials). To achieve this, the computer program arranges molecules on a kind of odor map, where flowery smells are located closer together than, for example, rotting ones. The study may significantly increase our understanding of how people perceive odors by quantitatively classifying odors. AI may be heralding a revolution in the study of this more mysterious human sense, as it has already done for the study of vision and language.
“The last time we digitized a human sense was a generation ago”, Alex Wiltschko, a neuroscientist and co-author of the paper, said. “These opportunities don’t come around that often”. Even though computers still can’t smell, this research is a significant step in the right direction. Wiltschko started working on this project at Google Research, and his start-up, Osmo, is now dedicated to it. “People have been trying to predict smell from the chemical structure for a long time”, Hiroaki Matsunami, a molecular biologist at Duke who studies olfaction and was not involved with the study, said. “This is the best at this point in order to do that task. In that sense, it’s a great advance”.
The only data accessible for a fragrance comes from human noses and brains, which are notoriously poor sources of data for machine-learning algorithms. Even small alterations to a molecule can turn a lovely, banana-scented substance into a compound that smells like vomit; strange changes to your nose and brain can turn coffee into sewage.
With the help of researchers in the flavor and fragrance industries, Wiltschko and his team set out to identify and curate a collection of about 5,000 molecules and the odor descriptions that went along with them (such as “alcoholic”, “fishy”, “smoky”, and so on). They then fed this data to an algorithm known as a graph neural network, which was able to represent the atoms and chemical bonds of each molecule in the form of an internal diagram. Given a molecule’s structure, the resulting program can forecast how it will smell using a mix of the existing odor labels.
Assessing the precision of those forecasts posed a different problem. A brand-new, independent group of individuals had to be trained to smell and categorize a brand-new set of molecules that the software had never studied. According to Joel Mainland, a neuroscientist at the Monell Chemical Senses Institute in Philadelphia who assisted with the study’s training, “People are really bad at [describing scents] when they walk off the street”, Joel Mainland, a neuroscientist at the Monell Chemical Senses Center in Philadelphia who helped conduct the training for the study, said. “If you train them for a couple of hours, they get pretty good, pretty fast”.
Participants were given various items, such as kombucha (“fermented”), a crayon (“waxy”), or a green-apple Jolly Rancher (“apple”), throughout five one-hour sessions to learn a reference point for each label. According to Emily Mayhew, a food scientist at Michigan State University and co-author of the study, participants then took a test in which they had to describe the smell of 20 common molecules (vanillin is vanilla-scented; carvone is minty), and they then retook the test to ensure their evaluations were accurate. Everyone who succeeded could help with algorithm validation.
The researchers asked participants to smell and describe all of the new molecules with different labels, each rated from zero to five (for example, a lemon might get a five for “citrus”, a two for “fruity”, and a zero for “smoky,” hypothetically). The new molecules were chosen by the researchers to be very different from the set used to train the program. The benchmark used to evaluate the machine was the sum of all those ratings.
“If you take two people and you ask them to describe a smell, they will often disagree”, Mainland said. But an average of several smell-trained people is “pretty stable”.
In general, the AI model “smelled” a little bit more accurately than the research participants. Sandeep Robert Datta, a neurobiologist at Harvard who did not conduct the research but serves as an informal advisor to Osmo, described the program as “a really powerful demonstration that some key aspects of our odor perception are shared”. A lemon may smell differently to different people, yet most people can agree that while an apple does not smell citrusy, both an orange and a lemon do.
The study’s map is another factor. Every molecule, and hence its odor, may be quantitatively represented in a space of mathematics known as a “principal odor map,” according to the authors. According to Wiltschko, it offers insight into the relationship between structure and smell as well as the way our brain categorizes odors. Floral scents are located in one area of the map, whereas meaty scents are located in another. Lavender is located closer to jasmine on the map than it is to a beefy aroma.
Datta warned against calling the odor map a principal rather than a perceptual tool. “It does a beautiful job of capturing the relationship between chemistry and perception”, he said.
Yet it doesn’t account for all the processes that take place as a molecule is converted into chemical signals, which are then converted into verbal descriptions of a smell, from receptors in our nose to the cerebral cortex in our brain. The map also differs from RGB (vision) values in that it does not list the fundamental elements necessary to create any particular fragrance, however, it does “suggest to us that RGB [for smell] is possible”.
He went on to say that the computer model’s perceptual odor map is an “extraordinarily important proof of concept” and offers vital details about how the brain allegedly organizes odors. For instance, Datta explained, you might believe that some types of smell—like citrus and smoky—are completely distinct. Yet, the odor map implies that even these dissimilar scents have connections.
The model is merely one of many developments required to digitize fragrance. The authors of the paper easily acknowledge that “it still lacks some of the important aspects of smell”, as Matsunami said. As most naturally occurring scents are the product of extremely complex combinations, their program is unable to anticipate how molecules will smell when combined. A smell’s intensity, as well as its quality, can vary depending on its quantity. For example, the molecule MMB, which is added to household cleaners and emits a nice smell in tiny amounts, contributes to the cat urine stench when it is present in high concentrations.
Given that people’s unique senses vary, it is unknown how well the software would perform in real-world scenarios, according to Datta, given that the model also predicts a smell only on average. Richard Doty, the director of the Smell and Taste Center at the University of Pennsylvania, who was not involved in the study, said that although the research is similar to the “Manhattan Project for categorizing odor qualities relative to physical, chemical parameters”, he is unsure of how much further the model can advance our understanding of smell given how complicated our noses are.
Wiltschko contends that additional study could address some of these issues and improve the map as a whole. For instance, the number of dimensions in the map is freely chosen to optimize the computer program; modifications to the training set of data may also enhance the model. Studying other components of our olfactory system, such as neurological routes to the brain or nose’s receptors, may also serve to shed light on how and at what stages the human body processes different odors. One day, a chemical sensor plus a set of computer programs that can translate the composition, concentration, and structure of molecules into a smell could realize digital olfaction.
It is somewhat astonishing that a computer model detached from the realities of human embodiment—a program that has no nose, olfactory bulb, or brain—can accurately forecast how something will smell even in the absence of good Smell-o-Vision. The research implicitly makes the case that knowledge of the brain is not necessary to comprehend smell perception, according to Datta. Using chatbots to explore the language network in the human brain or deep learning algorithms to fold proteins, the research highlights an emerging, AI-influenced body of knowledge. It is a comprehension that is more grounded in data than it is in worldly observation: prediction devoid of intuition.
This ground-breaking research into digitizing and quantifying smell may mark the beginning of the development of cutting-edge odor detection and reproduction technology. We might one day have devices that can “smell” items and substances in the environment if researchers can further improve computer models to properly forecast combinations of molecules, intensities, and variations across individuals.
Next-generation olfactory displays could be made by engineers using odor data and AI algorithms to generate complementary smells. They might be used in immersive virtual reality, where the realism is enhanced by synthetic scents of foods, flowers, or other objects. By enabling quick virtual testing and optimization, the digitizing scent could likewise change fields like food science and perfume creation. This discovery establishes a promising foundation for developments that might finally usher our chemical sense into the digital age, despite significant obstacles still standing in the way.