Can Machines Learn To Smell: Researchers May Have Cracked The Code

While CNNs provide vision to a machine, RNNs give machines the ability to draw insights out of speech and sound, haptics gives robots a sense of touch, researchers have now brought a new dimension in machines — the ability to smell.

While machines have been endowed with the sense of vision, speech, sound and touch thanks to the efforts of researchers for many decades, the concept of smell has eluded the machines. That is to say that humans have been unable to replicate this ability due to their own lack of understanding of olfactory functions.

Every aroma is a consequence of interactions between molecules. And, predicting the relationship between a molecule’s structure and its odour still remains to be a big challenge. But now researchers are trying to understand the interplay of molecules through quantitative structure-odour relationship (QSOR) modelling.

Developments Till Date

Till now the representation of molecular structures has been done with the help of graph theory. The way the symmetries and the connections can be defined as edges and vertices has been intuitive for further research into the three-dimensional structure of molecules. The same formulation was exploited a couple of years ago by the researchers at Google Brain in their quest to predict properties of molecules with machine learning. 

Download our Mobile App

Making use of the findings and moving forward, today, the researchers have leveraged graph neural networks (GNNs) to predict the olfactory properties of molecules.

How Graph Neural Networks Came In Handy?

By viewing atoms as nodes, and bonds as edges, we can interpret a molecule as a graph. These molecular graphs are encoded into a fixed-length vector. The fragments of this molecular graph give a hint of whether an atom or certain functional group exists or not.

This is where graph neural networks (GNNs) come into the picture. The researchers say that these GNNs are learnable permutation-invariant transformations on nodes and edges, which produce fixed-length vectors that are further processed by a fully-connected neural network.

However, translating the vector-node molecular representation to a graph is not as straightforward as it seems to be.

The translation into a graph happens as follows:

• Every node in the graph is first represented as a vector. For example, the node can be atomic charge
• Then, every node is made to pass on its current vector value to its neighbours. 
• An update function then generates a new vector value based on previous steps.
• This process is then repeated until all of the nodes are summarized into a single vector by summing or averaging
• This single vector represents a molecule, which is then passed into a fully connected network

The fully connected network gives an output that contains the prediction of molecules and their odour as provided by perfume experts.

Researchers have also found that the learned embeddings from graph neural networks capture  a meaningful odour space representation of the underlying relationship between structure and odour, as demonstrated by a strong performance on two challenging transfer learning tasks.

Future Direction

Understanding the functioning of olfactory sensory reception in itself is an exciting domain. Odour perception in humans is believed to be the result of the activation of 300-400 different types of olfactory receptors, expressed in millions of olfactory sensory neurons, embedded in a small patch of tissue called the olfactory epithelium. 

An aroma invokes many memories. Every fragrance is associated with memory. So it can be safely assumed that deciphering the inner workings of odour-memory association work can uncover the workings of memory itself.

From sniffer dogs in airports to wine connoisseurs, the sense of smell permeates into many real-world applications. Since every living and non-living thing can be melted down to the rubrics of chemistry, the use of machine learning for the same might unveil many intricacies of life.