Artificial Intelligence Learns to Smell

Did you ever think about how artificial intelligence will deal with smell in the future? A group of 22 teams of computer scientists came up with a set of algorithms able to forecast the scent of different molecules based on their chemical composition. With such algorithms you might be able to know exactly what people will smell in the future.

Compared to the sense of sight, the human sense of smell is a barley researched topic and a hard one to grasp for artificial intelligence. Smell is a deeply rooted human sense, million years ago it might have been even more important to us than vision. This changed drastically through evolution, our constantly developing distinct sight is supposed to be one big factor that helped us growing into the species we are at this present stage. Visual perception is well studied and easier to predict for artificial intelligence.

Leslie Vosshall, a researcher at the Rockefeller University in New York City, was the first one to work on predicting smell. Her team tested 49 volunteers and had them rate 476 smell samples of genuine odorants. The “smellers” where able to decide between 19 different categories containing “fish,” “garlic,” “sweet,” or “burnt.” Pleasantness and intensity of the smells where also taken into account. Resulting in a impressive database of over one million data points describing all the odor molecules tested in the study. Based on Leslie Vosshalls study, computational biologist Pablo Meyer set up the DREAM Olfaction Prediction Challenge.

With the chemical structure of the smell of molecules and their descriptors including atoms, their arrangement and geometry, in total two million data points where reached. This data was then used to train computer models in predicting smell, based on their chemical facts. In addition, a part of the data was implemented to test how exact the models forecast a person’s smell and in which category they would sort it.

The overall outcome of the study is that computers can now predict which of 19 words people will use to describe the tested set of odors. Avery Gilbert, a biological psychologist at Synesthetics and pioneer in the field of olofactation, sees the research critical. He thinks its not clear whether the same AI program would be able to tackle a wider set of categories. “If you had different descriptors, you might have had different models predict them best. So I’m not sure where that leaves us” Gilbert explained.

Summing up, AI is not able to know entirely how and what we smell yet, but maybe the outcome of this experiments shows us that there is still a lot to find out about the way the sense of smell can be used in new technologies and it will be a challenge for both, human scientist and artificial machines.

Source: Sciencemag. Image: Hugh Kretschmer

Welcome back!

We have noticed you are a frequent visitor to our website. Do you think we are doing a good job? Support us by becoming a member.