Unlocking the most undervalued sense
This winter, I attended MIT's winter course on deep learning (6.S191), and was transfixed by a session on Machine Learning for Olfaction, led by Dr. Alex Wiltschko from Google Brain.
It's astonishing how far ahead developments in computer vision and hearing are from their pungent counterpart. This is actually rather expected, since the sense of smell is just an all round underdog. According to Wired, a 2011 poll found that 53% of young adults would rather give up their sense of smell than give up their smartphones and computers. I wouldn't... not only because my phone is backed up, but also because I am rather fond of my particularly astute sense of smell (+ taste). That said, I had never considered the immense power of olfactory tech.
Smell basically IDs whatever object is emanating it, in a way not possible by vision alone. This opens the doors to a supreme array of applications, which were furiously bouncing around my head during the session.
Scent also seems to deeply associated with memory, and I sometimes smell things when recalling strong memories... 🍃The perfume industry is probably making key advancements concerning smell, presumably focussed on concocting nice-smelling, memorable, commercializable smells.
A quick aside on how smell works (in mammals... insects like ants apparently smell through their antennae _(ツ)_/¯) - smell is transmitted through receptor cells to the brain. I suppose this means the physical shape of a molecule activates a selection of electrical signals, which in turn stimulate various parts of the brain that triggers different smell-effects. During the session Dr. Wiltschko shared that electronic noses attempt to mimic this behavior by determining the molecular weight and structure of air borne particles.
He also shared some technical challenges of sense making from electronic noses, driven by the fact that molecules which are atomically and structurally similar can smell different (is smell even objective?), and those that are different can smell similar. Naturally, some smart folk have attempted to distill principles for determining smells (e.g. Ohloff's rule, Kraft's vetiver rule, Bajgrowicz and Broger's ambergris osmophore rule, Buchbauer's santols, Boelen's synthetic muguent), but it appears that general laws have yet to have been discovered.
And so, with our (my) somewhat foggy understanding of smell, we venture into creating computational models to make predictions from molecular fingerprints, most accurately by converting molecules into graphs for neural network analysis (GNNs).
Apparently, the representations generalize to achieve state-of-the-art on two major olfactory benchmark tests (DREAM Olfactory Challenge and Dravnieks).
Upon further research, I found that dogs have one of the highest accuracy rates when it comes to discriminating smells, and can be trained to detect diabetic emergencies, epilepsies, and even bladder cancer! Incredibly, dogs trained to detect bladder cancer (from pee) can generalize to detect other kinds of cancer as well 😮Good doggo!
This is undoubtedly the coolest thing I've come come across this winter, and I can't wait to dig deeper into the techniques and applications of machine smelling.
To be continued...
