How do you teach a computer to smell?
By Emre Ozer | Pragmatic
A friend buys you a bunch of flowers. The first thing you do is plunge your nose in and inhale. It’s an automatic – almost visceral – reaction. The gesture is simple, but getting your brain to make sense of the information it’s getting from your nose is, in fact, a highly complex operation.
Each odour is made of thousands of tiny molecules. When these molecules enter the nose – aided by a vigorous sniff – they travel to the olfactory cleft, where the nerves for smell are located. Here, the specialised cells – the olfactory sensory neurons – send messages to the brain, enabling it to register the smell and identify it as pleasant or unpleasant.
So, since it’s lacking both nasal cavity and olfactory sensory neurons, how do you teach a computer to smell?
Essentially, the answer is that you expose it to lots of data and get it to ‘learn’. But, once again, the simplicity of the statement belies the complexity of the operation.
Teaching a computer to smell
Earlier this month, my colleagues and I published a paper in Nature Communications detailing our research project – in collaboration with the University of Manchester, Unilever and Arm – on malodour classification with low-cost flexible electronics.
In other words, teaching a computer to smell. Initially, we targeted the detection of body malodour for application within the hygiene industry – specifically, personal deodorant.
The global deodorant business is a multi-billion-dollar operation, but current malodour measurement is surprisingly low-tech. It relies on an expert panel of human armpit assessors to sniff their subjects and rate odour quality, assigning a score to each test subject. This provides an output that is highly reproducible. But it’s also time consuming, costly and not particularly scalable.
To address these issues, we proposed a technology-based solution to automate the task, extending our work on the Innovate UK-funded PlasticArmpit project, which itself built on the success of PlasticARM – the world’s first fully functional non-silicon Arm processor.
A unique solution
In comparison to, say, detecting rotten apples, teaching a computer to assess body odour is a particularly complex task. As they ripen, apples follow a predictable aroma pathway, but the composition of our bodily secretions varies according to a range of factors including age, sex and diet.
Using an ‘electronic nose’ sensor array, a sensor readout interface and machine learning hardware based on low-cost flexible substrates, we developed a custom-designed classification system to objectively rate the funkiness of the test subjects’ personal aroma.
The decision-tree-based machine learning model is trained using data obtained from human armpit odour assessors, and returns a numerical value based on the strength of the odour. Thus, the returned value is equivalent to a human assessment.
Although we’re not the first to use e-nose functionality, our project is the world’s first to demonstrate a system integrating an e-nose sensor, sensor interface and a predictor hardware in low-cost flexible electronics.
These tiny devices can be embedded in ordinary fabric to create smart swatches, which are then used to swab a subject’s underarm. Once the swatch is swabbed, a malodour score is assigned and displayed directly on the swatch itself.
Developing a nose for the future
Beyond the initial scenario of ‘smelling’ body odour, our solution has potential application in a huge range of areas.
Added to food packaging, it could give a reliable indicator of food freshness, based on the presence – or otherwise – of odours that indicate deterioration. By displaying that information in a customer-friendly format, it could prevent unnecessarily short use-by dates, while maintaining customer confidence in the quality of the product.
In medical scenarios, it could be used to detect surgical site infections, currently developed by one in 241 patients in the US. The detection of changes in the volatile organic compounds emitted during the healing process is emerging as an effective method for monitoring wounds. Our solution could provide a low-cost and efficient way to alerts clinicians to early signs of deterioration, before more invasive treatment is required.
Another exciting possibility, which is, admittedly, further down the line, is the potential for detection of diseases, such as Parkinson’s. Recent studies have shown that the skin of people with early Parkinson’s produces a particular odour. Applying an e-nose could provide a quick, easy and scalable way to test patients before their motor symptoms are affected, enabling treatment to prevent the disease from spreading.
The potential applications are many and varied – but all exciting. While it’s difficult to predict exactly how the technology will evolve, it’s clear that the potential benefits of teaching a computer to smell are simply too compelling to ignore.