When AI learns to smell, humans can save 70 years of work

Image source: Generated by Unbounded AI

I don’t know how many people still remember Google Nose.

This funny project launched by Google on April Fool's Day in 2013 claims to have a smell database containing 15 million flavors. Users only need to enter keywords in the Google search box and click "smell" to smell them directly next to their computer. The smell of the object, such as the smell of a new car, the smell of a campfire, the smell of an Egyptian tomb (?), etc.

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

"Science" magazine in early September this year published a paper jointly published by multiple research teams including the startup Osmo (spin-off from Google) and the Monell Chemical Senses Center (Monell Chemical Senses Center), which stated, ** AI models can give machines a better “sense of smell” than humans**.

At first glance, this sounds incredible. After all, to the public, the sense of smell is a much more abstract existence than vision and hearing. The RGB color spectrum can describe the colors seen by the human eye, and the sounds heard by the human ear can also be converted into wavelengths of different frequencies, and even make people feel vibrations. However, only the sense of smell cannot be seen or touched, and it is even more difficult to describe with quantitative indicators. .

In other words, digitizing smell sounds impossible.

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

So how exactly is it done?

We know that odor is the response of the human olfactory system to certain specific molecules scattered in the air. After the odor molecules enter the nostrils, they will react with the olfactory cells above the nasal cavity (receptors), and the bioelectric waves generated will be transmitted to the brain through nerves, and then the smell will be recognized.

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

So faced with such a complex olfactory mechanism, the first thing the researchers did was to create a machine learning model—Message Passing Neural Network (MPNN).

Model diagram

This is a specific graph neural network (GNN), because 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 Used to learn complex odor features.

After the model is built, the next step is to feed it learning materials.

The researchers combined the Good Scents and Leffingwell & Associates (GS-LF) flavor and fragrance database and established a reference data set containing about 5,000 molecules as the basic training material. Each molecule can have multiple odor labels, such as fruity, Floral, cheesy, minty and more.

Some molecules in the GS-LF database

By taking the shape and structure of the molecule as data input, the model is able to output corresponding odor words that best describe a certain odor.

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

The experiment will eventually form the following high-dimensional olfactory map POM (partial):

This picture intuitively represents the perceptual distance of each smell. For example, there are large perceptual distances between floral, meaty and ethereal categories; but under each category The more specific smells included, such as lily (muguet), lavender (lavender) and jasmine (jasmine) under floral fragrance, have a closer perception distance.

The paper compared POM with Morgan fingerprint-based maps, which have been studied before, and found that the latter cannot yet reflect the above-mentioned perceptual distance:

In order to further verify the model training effect, the researchers then recruited 15 smell experts to compete with the model to see who could identify smells more accurately.

Each of the 15 experts needs to smell 400 odors. The researchers will give 55 odor adjectives and ask them to rate the 55 options on a scale of 1-5 for each odor to evaluate the extent to which each odor adjective is suitable. for this smell.

It was found that for 53% of the test molecules, the model performed better than the average of the panel members.

The researchers also classified the model's prediction results by odor descriptors and found that, except for musk, the model's prediction results for molecular odors were all within the error distribution of the human group, and outperformed the prediction results of 30 odor descriptors. Human group median:

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

Now we enter the most exciting stage of large-scale drawing of odor maps, and finally get the following picture:

You can understand the above coordinate diagram indicating the smell perception distance as an infinitely enlarged version of this diagram. The paper mentions that this map contains about 500,000 odor molecules, many of which have not even been discovered or synthesized (but can indeed be calculated).

To make a more intuitive comparison, if a trained human evaluator were to search for these smells, it would take about 70 years of continuous work 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 opinions, such as thinking that it can be used for quality control of factory sewage treatment, sniffing for explosives, drugs or corpses, etc.:

As a result, police dogs and search and rescue dogs may be off duty.

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

Some people are also very interested in the medical applications of this research result, such as the development of new treatments for anosmia, or the detection of diseases through smell, etc.:

There are also practitioners in the perfume industry who feel that this has helped them a lot, "It tells my colleagues when they wear too much cologne":

These predictions are actually not unreasonable. First of all, machines can indeed help humans solve the problem of sometimes inaccurate identification of smells - research shows that everyone has different perceptions of smells, and will trigger different reactions based on sensory and physiological signals, which are also affected by experience, expectations, and personality. or the influence of situational factors.

And smell is sometimes very important to people.

Needless to say, bad smells, some harmful gases may also be harmful to health. At this time, it would be great if machines could replace certain occupations to help humans or animals work.

For other professions where scent can bring benefits, such as perfumers, chefs, designers, artists and architects, etc., there is also a need to prepare more functional scents. Some occasions use scents in the environment. For example, the Sloan-Kettering Cancer Center in New York disperses vanilla oil in the air to reduce patients' claustrophobia during magnetic resonance imaging (MRI) tests; the Chicago Board of Trade also disperses specific scents. To reduce noise decibels on the trading floor.

Other studies have shown that most of humans' odor-related memories come from the first ten years of infancy and early childhood, while memories generated by language and vision are usually produced between the ages of 10 and 30. This partly explains why smells can evoke distant memories, and memories evoked through smells are often more emotionally charged than memories evoked by sight or hearing.

Therefore, the connection between smell and human beings is still very close, but we are not easily aware of it in many cases.

Netizens’ conjectures were also verified by one of the authors of the paper, Alex Wiltschko from Osmo Company. He wrote in an article posted on Osmo’s official website,

“Smell mapping is the basis for our larger goals. If a functional system that replicates our nose or a dog’s nose can be developed, we can detect diseases early; artificial intelligence will also help doctors find more likely to be detected in the clinic. to develop successful drugs and better assist synthetic chemists and master perfumers in their work...Our future work goal is to lay a solid scientific and commercial foundation for improving human health and happiness."

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

For example, it is impossible to reflect the intensity of a molecule's odor, and can only predict what it smells like; only the odor of a single molecule is predicted, but in real life it is more of a mixed odor; and even if all abilities are achieved, the odor cannot be predicted. The replication and restoration will also be a big challenge and so on.

Finally, having said so much, one netizen’s comment was quite simple, “I think this will make wine tasting less fun”: