Algorithm sniffs solution to problem of smell

Smell is a messy business. In 2004, American author Chandler Burr described at length just how complex the subject is in a surprisingly gripping account of the work of biophysicist Luca Turin, and his long-running and often acrimonious battle with the perfume industry.

Burr’s book, The Emperor of Scent, centres on radically different theories of how odour detection actually works on a molecular level.

Turin – who is still researching and publishing in the field – contends that smell is governed by molecular vibration. For example, if a given aromatic molecule has a hydrogen atom replaced by a heavier deuterium atom, the human nose should detect it as having a different smell, even though the two molecules have precisely the same shape.

Results and interpretations around the hypothesis continue to differ and, often, arguments about the implications rage. (The perfume business is nothing if not bitchy.) The central take-home message, however, is clear: smell is a fearfully convoluted process.

This is a matter cheerfully accepted by computer scientists Nabil Imam and Thomas Cleland – from Intel Corporation and Cornell University, US, respectively – in a new paper published in the journal Nature Machine Intelligence.

In the paper the pair announce the creation of a neural algorithm that can learn to identify odour samples, and, what’s more, can do so even when the sample is buried inside noise.

In other words, the researchers have invented an algorithm – and, by extension, the machine that houses it – that can smell.

Of course, that term needs to be clarified. Smell for humans is tied up with a wide range of environmental, survival, and emotional cues. Roses smell sweet; hot chips induce salivation; rotting flesh prompts a flight response. 

Imam and Cleland’s algorithm does none of that. It simply identifies a specific molecule, which, should it come into contact with a mammalian main olfactory bulb (MOB), would send signals to the brain that would be interpreted as a smell.

The researchers modelled their algorithm on the neural circuitry of the external layer of the MOB, building a network of excitatory and inhibitory units embedded in Intel’s newly developed high-tech neuromorphic chips, called Loihi.

They trained it using a large data set published in 2013 by another US scientist known for his work on artificial noses, Alexander Vergara from the University of California San Diego. He and colleagues suspended 72 metal oxide chemosensors inside a wind tunnel, and used them to record 18,000 time-limited measurements of the molecular components of a wide range of odourants.

The results are these days often used as a standard baseline for scientists in the field. 

In their paper, Imam and Cleland report that their algorithm is designed to learn from “one-shot” input – the equivalent of a single sniff. Once learned, the program “remembers” the input, storing it away. It is capable of learning to identify multiple odourants, building its own data bank.

The first test was made using toluene – the aromatic hydrocarbon that gives pain-thinners their distinctive smell. The attempt was successful, and the algorithm was able to subsequently detect it even when it was buried inside a lot of added irrelevant signal.

Subsequent successful tests were carried out using ammonia, acetone, carbon monoxide and methane.

The ability of the system to find aromatics inside messy data is important, because in the physical world smells are almost never pure and isolated – nor stable over even small timeframes.

The researchers, however, state that their creation is a step forward, but may well not work – in its current state, at least – in real environments.

“While both biological and artificial olfaction systems recognise chemical analytes based on activity patterns across arrays of weakly specific chemosensors, mammalian olfaction demonstrates levels of performance in signal restoration and identification currently unmatched by artificial systems,” they write.

“Indeed, the underlying identification problem is deceptively difficult.”

In other words, as Luca Turin and the boffins at Chanel and Prada realised many years ago, smell is anything but simple.