The computer that can smell - how Intel is revolutionizing AI

Close-up of the Loihi, Intel's latest neuromorphic research chip, from Intel Newsroom.

Out of sight, hearing, touch, smell, and taste, which of the five senses seems most difficult to recreate in AI? To me, scent seems among the most difficult, if not the most. I mean, imagine a computer that can smell. Can we create electronic noses now too? Well, maybe we can (kinda), because researchers from Intel and Cornell University recently trained a neuromorphic chip to recognize the scents of 10 different chemicals. The goal? In the future, this technology might enable "robotic noses" to detect weapons, poisons, diseases, and more.

Mimicking biology... with silicon!

First, teams from Intel and Cornell examined what happens in a mammal's brains when it smells something. Your nose contains about 450 different olfactory receptors. These can be activated by airborne odour particles to send signals to your brain, where electrical pulses activate neurons and generate the sensation of the smell corresponding to the particle.

How were the researchers able to mimic this biological pathway using circuitry? How did they achieve the intense processing power required to "smell"? By using Intel's Loihi, they designed an algorithm modelled as closely as they can to the brain's olfactory circuit. Loihi is Intel’s neuromorphic computing chip. If you haven't heard of neuromorphics yet, I encourage you to read more on the topic - it's a very revolutionary technology in the realm of AI. Neuromorphic computing applies the principles of computation found in your human brain, to computer architectures. So the Loihi creators are essentially trying to fit as much of your brain's functions as they could in a 14-nm, 130,000-"neuron", 130 million-"synapse" heavy system. Yes, you read that right. They are trying to recreate neurons and synapses, terms you'd be very familiar with if you've studied biology, inside a 14 nm chip! That's pretty crazy already, but it doesn't stop there. 

After design comes training in ML and AI. Researchers used a dataset consisting of the activity of 72 chemical sensors in response to 10 odours. They recorded the response of these 72 chemical sensors in a wind tunnel as they circulated ten different scents (including methane, ammonia, acetone, etc...) in the tunnel. The sensors' responses were sent to the Loihi, where silicon circuits mimicked the circuitry of the brain dedicated to the sense of smell, thus drawing neural representations of each of these smells. This was their attempt at recreating how the brain assigns scents to different patterns of electrical signals.  The chip was able to learn neural representations of each of the 10 smells, and was able to identify them, even in the presence of strong background noise in the form of other scents. This is how the Loihi tried to recreate the same biological "circuitry" in our brain that makes us jump at the smell of something burning, or gag when we enter a less-than-pleasant restroom...

Why this is revolutionary 

Firstly, the fact that it's even possible to mimic the pathways in our brain with silicon and copper just blows my mind. Ponder that for a moment, and let it blow your mind. Secondly, the smoke and carbon monoxide detectors at your home use sensors to detect odours, yes, but they can't distinguish between them. They set off an alarm when they detect harmful molecules but are unable to categorize them in intelligent ways. Not to mention, they are only able to detect smoke and carbon monoxide molecules, not particularly distinguish them due to "scent." The ability to categorize is key here, because our ability as humans to categorize many different stimuli form the basis of our decision making skills. If an artificially intelligent robot is going to learn to make decisions like us, it's going to have to get really good at classification (categorizing) first.

Intel is still faaar from replicating the brain's entire olfactory system on a chip, however. Let's not disgrace the beauty and complexity of the human body, and especially our brains, by disregarding this fact. This is, however, a leap in AI. The technology could be used for medical diagnoses where conditions, pathological or not, can be differentiated by scent. This technology can also allow robots to identify hazardous substances in airports and other high-security areas.

There are many challenges in olfactory sensing, however. Even the human nose can't differentiate between similar scents sometimes. How then do we mimic these imperfect pathways using electronic circuitry? How do we perfect the system? A strawberry, blueberry and banana all induce very similar neural activity patterns in the brain, according to the research team at Intel. According to them, a robot may get confused when it smells a strawberry from Italy and one from California. These two strawberries may have different aromas, but are both strawberries, and thus must be grouped into a  common category. The team at Intel recognizes that they must solve these challenges in order for this product to be able to perform in the real world.