Will my dog have a job?
Over the next few weeks of our blog we’ll be exploring how data science and AI play a role in the world around us. Lately I’ve been considering what kinds of tasks in my daily life might be replaced by AI systems in the future, and what that might look like.
In my spare time I’m involved in training dogs to find missing people. I’m constantly amazed at the scent abilities of dogs - from discriminating between individuals to following week-old scent trails. In our training we try and set up complex scent “puzzles” that the dogs need to work through. Us humans imagine these scent puzzles to be complex, but the dogs generally breeze through with ease. There is a whole world of scent in our physical environment that isn’t easily discernible by us humans, but can be tapped into by and through our dogs.
Various industries have harnessed dog’s incredible scent capabilities and use them to detect land mines, drugs, explosives, corpses, firearms, illegally poached and trafficked wildlife, rare and endangered animals (there is a tortoise detection dog working on Table Mountain), various types of cancer, tuberculosis and even Parkinson’s disease. The possibilities for detection are almost endless and in some cases dog detection fares better than lab tests in terms of accuracy.
But finding suitable dogs is tough - there is currently a massive shortage of dogs for detection work and trained dogs can fetch prices as high as $25 000. Dogs also take time to train, they get tired and they can be prone to distractions. Despite the accuracy of dog detection, you don’t find dogs in hospitals waiting around to detect cancer. And so there is a very clear need to develop artificial olfaction solutions.
Artificial olfaction, the automated simulation of the sense of smell, is based on electronic systems (called electronic noses) comprised of an array of sensors of some type, the electronics to interrogate those sensors and produce digital signals and data processing and user interface software. Humans and animals don’t identify specific molecules within odours, rather we recognise a smell based on a response pattern. And so the crucial part of detecting scent is pattern analysis of the response to an odourant. This involves processing of multivariate data - a realm with which we are more comfortable - and may include feature extraction, feature selection, classification, regression and validation. Electronic noses are first trained on a target odour and then used to detect that odour in other or future samples.
There are several electronic noses (or e-noses) on the market but the devices are bulky and expensive. Currently most applications focus on detection and identification of various odours, particularly for military applications and disease detection. There is certainly huge demand for smaller, more robust and cost effective e-noses. I wonder how soon it will be before we have a device that is capable of following a scent trail over tough terrain to find a missing person. We might be relying on our furry friends for quite a few more decades.