Can You Teach a Chip to Smell?
Researchers at Intel and Cornell University are bringing the "science of smell" down to the silicon level.
On March 16, 2020, Intel announced that their researchers, along with researchers at Cornell University proved that chips can smell—with the right amount of training, that is.
These researchers demonstrated the ability of a self-learning neuromorphic chip to recognize hazardous chemicals, even while surrounded by significant occlusion and noise. When the chip, Loihi, was tested with single samples of an odor, it effectively "learned" the odor without disturbing the memory of previously learned scents.
Nabil Imam, a senior research scientist in Intel Labs’ neuromorphic computing group, holds a Loihi test chip, which he uses with a research team from Cornell University to build mathematical algorithms to mimic what happens in the brain’s olfactory system. Image used courtesy of Walden Kirsch and Intel
Overall, the Loihi was able to learn 10 different odors. The performance of the Loihi was found to be superior compared to traditional methods, such as a deep learning solution that required 3,000 times more training per class to have the same level of classification accuracy.
In September 2017, Intel released the Loihi, their neuromorphic research test chip with digital circuits that mimic the brain’s basic mechanics. The research team did this to speed up machine learning while making the process more efficient with less computing power.
Intel’s Loihi is a self-learning, neuromorphic chip designed to help researchers make significant progress at the intersection of neuroscience and artificial intelligence. Image used courtesy of Intel
The Loihi chip features a fully asynchronous neuromorphic many-core mesh that supports various neural network topologies and gives each neuron the capability of communicating with thousands of other neurons.
Each core itself has a learning engine that can be programmed to adapt network parameters during operation and can support supervised, unsupervised, reinforcement, and other learning paradigms found in machine learning.
Whereas the human brain has over 86 billion neurons and 100 trillion synapses, Loihi has 130,000 neurons and 130 million synapses fabricated with Intel’s 14 nm process technology.
How Did Loihi Learn to "Smell?"
So, how did researchers achieve this feat? They started with a neural algorithm that they based on the architecture and dynamics of the human brain’s olfactory circuit. Next, they trained Loihi to recognize the scent of 10 hazardous chemicals.
Nabil Imam, a senior research scientist in Intel Labs’ neuromorphic computing group, and his team then compiled a dataset comprised of the activity of 72 chemical sensors in response to these 10 scents, including acetone, ammonia, and methane.
The chip was designed to mimic the ways the human brain distinguishes smells. Screenshot used courtesy of Intel
The sensor responses to the individual scents were transmitted to Loihi where the circuitry of the brain responsible for the sense of smell was mimicked by silicon circuits. A significant feat of Loihi is the ability to distinguish the difference between smells, even with strong background interferents. For comparison, smoke and carbon monoxide detectors at home can detect odors, but cannot distinguish between them.
The Future of “Electronic Nose Systems”
According to Imam, the chemical-sensing community has been searching for an inexpensive, reliable, and fast-responding chemosensory processing system like Loihi. Such a system is also known as an “electronic nose system.”
Some uses of these systems can include:
- Equipping robots with neuromorphic chips for environmental monitoring and hazardous materials detection; this can allow researchers to know exactly what gaseous substances are being released into the atmosphere.
- Controlling the air quality in factories
- Diagnosing medical conditions in cases where diseases emit particular odors (similar to how dogs can smell certain diseases in humans)
- Identifying hazardous substances in airport security lines such as bombs or narcotics
Furthermore, Imam hopes to “generalize this approach to a wider range of problems” to understand relationships between observed objects and solving abstract problems, like planning and decision-making.
Being able to translate the brain’s olfactory system into digital circuitry helps researchers understand how neural circuits solve complex computational problems and provide insight into designing “efficient and robust machine intelligence,” Imam states.
The Challenges in Olfactory Signal Recognition
The significant progress made by the teams at Intel and Cornell University does not come without several challenges to be addressed in future designs.
Much like how humans may have trouble distinguishing the smells of fruits like blueberries or bananas because of the similarity in neural activity patterns in the brain, neuromorphic systems can face similar problems—especially when attempting to classify distinct smells in common categories.
A close-up image of Loihi. Intel announced that 64 Loihi chips will make up a new neuromorphic system called Pohoiki Beach. Image used courtesy of Intel
Imam believes that within the next couple of years, these challenges in olfactory signal recognition can be solved as the technology moves from the lab to real-world scenarios. Intel’s vision of the future puts neuromorphic computing at the forefront of solving complex world problems as the flow of intelligence and decision-making becomes accessible and accelerated.
In hindsight, would the "science of smell" be useful in any of your past projects? Why or why not? Share your thoughts in the comments below.