Machine learning has paved the way for voice and facial recognition, but researchers are still struggling to quantify the elusive and sometimes ambiguous sense of smell.
Machine olfaction is the automated simulation of the sense of smell, an emerging field that uses robots and other automated systems to analyze molecules in the air. As with vision and hearing, machine learning is key to digitizing smell, because machines can learn to map the molecular structures that create the smell and then translate that into a text description, according to Ambuj Tewari, a machine learning expert and statistics professor at the University of Michigan.
“The machine learning model learns words that humans tend to use, such as 'sweet' or 'dessert', to describe the experience they have when encountering a particular odor-causing compound, such as vanillin,” Tewari wrote in The Conversation.
Mechanical olfactory devices can also capture human odors and use them as biometric templates. In 2022, scientists at Kyushu University in Japan developed a sensor called an “artificial nose” that can biometrically identify a person from the smell of their breath with an average accuracy of more than 97%.
Some companies are already betting on this new field: Ozmo, a machine-olfaction startup, received a $3.5 million grant from the Bill & Melinda Gates Foundation last year to develop its AI-powered scent platform, and in early 2023 the company also received $60 million in Series A funding led by Lux Capital and Google Ventures.
The Google Research spinout wants to create a “smell map” that predicts a molecule's smell based on its structure. The platform will be used to create compounds that repel, attract or even kill disease-carrying insects like mosquitoes.
“Osmo's science has revealed an incredible connection between insect and human smells: our odor maps predict how molecules will smell not only to insects but also to humans,” says Osmo CEO Alex Wiltschko.
But quantifying smell is a difficult task, and not just because smell is hard to describe. The internet has a vast amount of audio, image, and video content that machine learning can use to train recognition systems. Tewari says machine olfaction has long faced a data shortage problem, and without datasets, researchers have struggled to train powerful machine learning models.
The breakthrough came with the 2015 DREAM Olfactory Prediction Challenge, which invited research teams from around the world to submit machine learning models. A research project called the Pyrfume Project has led to the release of many more datasets.
A research team led by Osmo and the Monell Chemical Senses Center at the University City Science Center campus in Philadelphia has finally been able to create an AI model that produces remarkable results in machine olfaction, paving the way for the prediction and digitization of smells.
The model can predict odor signatures based on a molecule's structure. It's based on a type of deep learning called a graph neural network and was trained on a dataset of 5,000 known odorants. The study, published in Science in September 2023, found that the AI outperformed human evaluations for more than half (53%) of the molecules tested, according to Neuroscience News.
Article Topics
Biometric Templates | Biometric Research | Body Odor | Machine Learning | Smell | Osmo
