When AI learns to smell, humans can work 70 years less

Author|Li Hezi

I don't know how many people remember Google Nose.

This funny project launched by Google on April Fool's Day 2013 claims to have a database of 15 million smells, users only need to enter keywords in the Google search box, click "smell", and directly smell the smell of the object, such as the smell of a new car, the smell of a campfire, and the smell of an Egyptian tomb (?). And so on.

It is this outrageous but brainy joke of ten years ago that is being partially turned into reality by its inventor.

In early September, Science published a paper by a team of researchers including the startup Osmo (spun off from Google) and the Monell Chemical Senses Center, arguing that AI models could give machines a better sense of smell than humans.

At first glance, this is simply incredible, after all, for the masses, the sense of smell is a much more abstract existence than sight and hearing. RGB color map can describe the color seen by the human eye, and the sound heard by the human ear can also be converted into wavelengths of different frequencies, and even make people feel vibration, but only the sense of smell, can not be seen and touched, and it is more difficult to describe it with quantitative indicators.

In other words, digitizing the smell sounds impossible.

The core of the researchers in this paper is to try to create a high-dimensional map of human olfaction, or POM (Principle Odor Map), that can faithfully reflect the characteristics of odor.

So how exactly do you do it?

We know that smell is the induction of certain molecules dispersed in the air by the human olfactory system. After the odor molecules enter the nostrils, they react with the olfactory cells (receptors) above the nasal cavity, and the bioelectric waves generated are transmitted to the brain through the nerves, and then recognize the taste.

The composition of odor is actually much more complex than color and sound, there are millions of different types, each smell is composed of hundreds of chemical molecules, their properties are different. Correspondingly, humans have about 400 functional olfactory receptors, far more than the 4 we use for vision and about 40 for taste.

So faced with such a complex olfactory mechanism, the first thing the researchers did was create a machine learning model, the messaging neural network (MPNN).

Model diagram

This is a specific graph neural network (GNN), because the graph neural network is a deep learning method based on graph structure, which introduces traditional graph analysis and provides a method for extracting features from irregular data, so it is also very suitable for learning complex odor features.

Once the model is set, the next step is to feed it learning materials.

The researchers combined the Good Scents and Leffingwell & Associates (GS-LF) flavor and fragrance database to build a reference dataset of about 5,000 molecules as a training base, each molecule can have multiple odor labels, such as if flavor, floral, cheese and mint.

Some molecules in the GS-LF database

By using the shape and structure of the molecule as a data input, the model outputs the corresponding odor word that best describes a certain odor.

In order to make the training results more accurate, the researchers also used various methods to optimize the model parameters. For example, the GS-LF flavor and fragrance database is divided into training set and test set according to the ratio of 8:2, and the training set is further divided into five cross-validation subsets; and using Bayesian optimization algorithms to cross-validate data and optimize the hyperparameters of GNN models.

The experiment will result in the following olfactory high-dimensional map POM (local):

This figure visually represents the perceived distance of each odor, such as floral, meaty and ethereal; But the more specific scents included under each category, such as muguet, lavender and jasmine, are more closely perceived.

The paper compares POM with Morgan fingerprint-based maps based on Morgan fingerprints, which has precedent and finds that the latter does not yet reflect the perceived distance:

In order to further verify the effect of model training, the researchers then recruited 15 odor experts to compete with the model to see who recognized odors more accurately.

Each of the 15 experts was required to smell 400 odors, and the researchers were given 55 odor adjectives and asked them to rate each odor on a scale of 1-5 for each of the 55 options, and how well each odor adjective was appropriate for the smell.

It was found that for 53 percent of the molecules tested, the model outperformed the average of the panelists.

The researchers also classified the prediction results of the model by odor descriptors, and found that except for musk, the prediction results of the model for molecular odors were in the error distribution of the human group, and the prediction results of 30 odor descriptors were better than the median of the human group:

Subsequently, the researchers also repeatedly verified the performance of the model and obtained a relatively stable molecular structure-odor relationship.

Here we move on to the most exciting large-scale drawing of the odor map, and finally get the following picture:

You can understand the above coordinate plot representing the perceived distance of odor as an infinite enlarged version of this graph. The paper notes that the map contains about 500,000 odor molecules, many of which have not even been discovered or synthesized (but can be calculated).

To make a more intuitive comparison, if a trained human evaluator were asked to look for these odors, it would take about 70 years to collect them all.

It seems that this paper has really accomplished a big thing. At this time, some netizens asked, why does the machine need to smell?

Others have also given their own insights, such as the idea that it can be used for quality control of factory wastewater treatment, sniffing explosives, drugs or corpses, and so on:

As a result, police dogs and search and rescue dogs may have to leave work

Some people hope that a good deodorant can be developed based on this, because people will emit a bad smell after doing a lot of aerobic exercise such as running or lifting weights:

There are also people who are interested in the medical application of this research result, such as the development of new treatments for olfactory loss, or the detection of diseases through smell:

There are also those in the perfume industry who feel that this has helped them a lot, "let it tell my colleagues when they sprayed too much cologne":

These predictions are in fact justified. First, machines can indeed help humans solve the problem of sometimes inaccurate identification of smells – studies have shown that everyone perceives smells differently, elicit different responses based on sensory and physiological signals, which are also influenced by experience, expectations, personality or situational factors.

And smell is sometimes very important to people.

Needless to say, some harmful gases may also be harmful to health, so it would be great if machines could replace certain occupations to help humans or animals.

For other professions that can benefit from scents, such as perfumers, chefs, designers, artists, and architects, there is also a need to formulate more functional scents. Some occasions apply odors to the environment, such as the Sloan-Kettering Cancer Center in New York, which spreads vanilla oil in the air to reduce claustrophobia for magnetic resonance imaging (MRI) tests; The Chicago Board of Trade also spreads a specific fragrance to reduce the noise decibels on the trading floor.

Studies have also shown that most human odor-related memories come from infancy and the first decade of early childhood, while memories produced by language and vision are usually produced between the ages of 10 and 30. This partly explains that smells can evoke distant memories, and that memories caused by smells are often more emotional than those caused by sight or hearing.

So smell and human connection are still very close, but we are not easy to detect in many cases.

The conjecture was also verified by one of the authors of the paper, Alex Wiltschko of Osmo. In an article posted on Osmo's official website, he wrote,

"The scent map is fundamental to our ambitions. If we can develop functional systems that replicate our nose or the nose of a dog, we can detect diseases early; AI will also help doctors find drugs that are more likely to be clinically successful, and better help synthetic chemists and master perfumers with their work... Our future work aims to lay a solid scientific and commercial foundation for improving human health and well-being. ”

However, he also said that the paper still has many shortcomings.

For example, it cannot reflect the intensity of the molecular odor, and can only predict what it smells; Only individual molecular odors are predicted, but in real life they are more mixed odors; And even if all the abilities are achieved, the replication and reduction of odors will be a great challenge, and so on.

Finally, having said all this, there is a netizen's comment that is very simple, "I think this will make wine tasting lose its fun":