Sunday, July 21, 2024

Giving computers a sense of smell

Over a century ago, Alexander Graham Bell encouraged National Geographic readers to undertake a bold endeavor: to establish a new science focused on smell. He noted that sciences based on sound and light already existed, but there was no equivalent for odor. Bell urged his readers to find a way to “measure a smell.”

Today, smartphones boast impressive capabilities derived from the sciences of sound and light, like voice assistants and facial recognition. However, the science of odor has lagged behind—until now. Advances in machine olfaction, or “digitized smell,” are finally addressing Bell’s challenge.

The complexity of human smell, involving about 400 types of receptor cells in the nose compared to the rods and three types of cones in the retina for vision, presents a significant hurdle for machine olfaction.

Machine olfaction begins with sensors that detect and identify airborne molecules, akin to the receptors in the human nose. However, to be useful, these systems must understand what these molecules smell like to humans, necessitating the use of machine learning.

Machine learning, particularly deep learning, has driven innovations such as voice assistants and facial recognition. It is crucial for digitizing smells because it can map the molecular structure of odor-causing compounds to textual odor descriptors. For instance, a machine learning model can learn that “sweet” and “dessert” describe the smell of vanillin.

Machine learning requires large datasets. Unlike audio, image, and video data readily available on the web, olfaction data has been scarce. This is partly because people find it harder to describe smells than sights and sounds, limiting the available data for training machine learning models.

The landscape changed in 2015 with the DREAM Olfaction Prediction Challenge, which provided data from biologists Andreas Keller and Leslie Vosshall. Teams worldwide were invited to develop models predicting odor labels like “sweet” or “fruit” based on molecular structure. The winning models, published in Science in 2017, utilized a machine learning technique called random forest.

As a machine learning researcher interested in chemistry and psychiatry, the DREAM challenge intrigued me. My connection to olfaction is personal, stemming from my roots in Kannauj, India’s perfume capital, and my father’s career as a chemist.

Progress in machine olfaction accelerated after the DREAM challenge. The COVID-19 pandemic, which brought attention to smell blindness (anosmia), and the Pyrfume Project, which released larger datasets, further spurred research.

By 2019, datasets grew from fewer than 500 molecules in the DREAM challenge to about 5,000. Google Research’s team, led by Alexander Wiltschko, applied deep learning through graph neural networks, achieving state-of-the-art results. Wiltschko, now CEO of Osmo, aims to give computers a sense of smell. His team created a “principal odor map,” positioning perceptually similar odors close together, despite challenges in olfactory perception due to small structural changes in molecules.

Advances in understanding smell have promising applications, including personalized fragrances, improved insect repellents, novel chemical sensors, early disease detection, and enhanced augmented reality experiences. The future of machine olfaction is promising and, undou btedly, smells good.

Leave a Reply