Google's Smell AI Can Predict Scents & Repel Mosquitoes
Principle Odor Map, a mapping of scents based on molecular structure, pushes new AI research into scent and posits use in mosquito repellant example.
Created on September 6|Last edited on September 6
Comment
In a blog post released today on the Google AI blog, researchers show off new smell-related AI research with Principle Odor Map (POM), a mapping of molecule scent data.
This research is an evolution of previous research done in 2019 which set the early stages for AI smell analysis using a new graph neural network model.
Visual and auditory mediums have always taken the front stage when it comes to sensory-based AI projects, but the medium of smell, the olfactory sense, is rarely considered. Perhaps, like a taste or touch AI, the utility is not immediately apparent, but this new research makes the case that there is certainly a space for AI models dealing with smells.
Understanding scent with AI and POM
One of the big pieces to highlight in this new research on display today is the development of a scent map, POM, which can represent smells in relation to each other similar to how colors are shown on a color map or wheel. POM, however, maps smells within a high-dimensional space based on the vector representations present in the smelling model's embedding space.
To evaluate the model, molecules were paired with the smells they produce, as determined by a team of trained panelists. Despite the training and carefully crafted selection of smell labels, inconsistencies were found among individuals on most labeling tasks. Smell is often very subjective, so this comes as no surprise.
Because of the individualized perception of smell, the model was evaluated against the decision consensus across all participants. It was found that the model could not just accurately predict the best scent descriptors for any given molecule but, compared to any individual in the scent judging team, it would match the consensus much closer.
Predicting new scents & repelling mosquitos
Because of the way this model can identify and categorize scents, it's capable of predicting what unseen molecules might smell like. One of the key things the researchers went to do with this capable AI was trying and find new scents which could potentially repel mosquitos.
The original model was modified with datasets of mosquito-molecule interactions and was able to determine the number of molecules which work to repel mosquitos even better than the standard-use insect repellant molecules DEET and Picaridin. Research like this could be crucial in preventing insect-spread disease by potentially lowering material and production costs for repellants with new molecules.
Find out more
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.