Researchers train artificial intelligence to recognize smells

A team of Google Brain researchers published a new study on arXiv in which they explain how they train artificial intelligence software to recognize smells. They first created a set of more than 5,000 molecules and then labeled these molecules with descriptions that identified the type of odor.

The researchers used a special artificial intelligence called graphical neural network (GNN) so that these molecules were associated with their descriptions based on their structures. This is not a software that can be compared to the sensitivity of the human sense of smell, because the latter is very difficult to define. For example, there are scents that can appear in one way for one person and in another.

Moreover, some molecules sometimes have the same atoms and the same bonds, but they are arranged as mirror images: these molecules, usually recognized by the same software as practically the same, can have completely different smells. And this without mentioning the fragrances that are the result of the fragrances combined.

Despite these undeniable difficulties, Google researchers think that this is an important first step, a step that can also be useful in the fields of chemistry, sensory neuroscience and the production of synthetic fragrances.

This is not the first team of researchers to attempt to mimic or imitate the characteristics of an olfactory system based on artificial intelligence. For example, a team of scientists from the Barbican Centre in London used machine learning techniques to “recreate” the smell of an extinct flower. In addition, IBM is conducting experiments to create new fragrances generated by artificial intelligence.


See also:

https://arxiv.org/abs/1910.10685

Image source:

https://o.aolcdn.com/images/dims?quality=85&image_uri=https%3A%2F%2Fo.aolcdn.com%2Fimages%2Fdims%3Fcrop%3D6000%252C3748%252C0%252C0%26quality%3D85%26format%3Djpg%26resize%3D1600%252C1000%26image_uri%3Dhttps%253A%252F%252Fs.yimg.com%252Fos%252Fcreatr-images%252F2019-10%252F1e49c110-f699-11e9-bb7b-04ecdc43f178%26client%3Da1acac3e1b3290917d92%26signature%3Daed5fe9f54b87d1f9801719d4003ea6ec92d25ef&client=amp-
blogside-v2&signature=5656ac71ea51a5dcd0e13bbf51aae68a548bc16c

Leave a Reply

Your email address will not be published. Required fields are marked *