fbpx

AI that mimics mammals smells

“Using the mammalian brain as a blueprint, scientists designed an artificial neural network that can keep learning new aromas without forgetting others.”

Original Source: https://www.sciencenews.org/article/ai-mimics-how-mammals-smell-superior-recognizing-scents

In identifying scents, a “neuromorphic” artificial intelligence beats other AI by more than a nose.

The new AI learns to recognise smells more efficiently and reliably than any other algorithm. And unlike other AIs, this system can keep learning new aromas without forgetting others, researchers report on Nature Machine Intelligence on March 16. The key to the success of the programme is its neuromorphic structure, which is more like neural circuitry in mammalian brains than other AI designs.

This kind of algorithm, which excels in detecting weak signals in the middle of background noise and continuously learning at work, could one day be used for monitoring air quality, detecting toxic waste or for medical diagnosis.

The new AI is an artificial neural network, composed of many computing elements that mimic nerve cells to process fragrance information (SN: 5/2/19). The AI “sniffs” by means of electrical voltage readings from chemical sensors in a wind tunnel that were exposed to feathers of different scents, such as methane or ammonia. When the AI smells new, it triggers a cascade of electrical activity between its nerve cells, or neurons, which the system remembers and can recognise in the future.

Like the olfactory system in the mammal brain, some of the AI’s neurons are designed to react to chemical sensor inputs by emitting differently timed pulses. Other neurons learn to recognise patterns in the blips that make up the electrical signature of the odour.

This brain-inspired setup puts the neuromorphic AI at the forefront of learning new smells more than the traditional artificial neural network, which begins as a uniform web of identical, blank slate neurons. If a neuromorphic neural network is like a sports team whose players have assigned positions and know the rules of the game, a normal neural network is initially like a bunch of random newbies.

As a result, the neuromorphic system is a faster, more nimbler study. Just as a sports team may need to watch a game only once to understand the strategy and implement it in new situations, the neuromorphic AI can sniff a single sample of a new smell in order to recognise the smell in the future, even in the midst of other unknown smells.

On the other hand, a bunch of beginners may need to watch a play many times to rehabilitate choreography — and still struggle to adapt it to future game-play scenarios. Similarly, a standard AI must study a single fragrance sample many times, and may still not recognise it when the fragrance is mixed with other odours.

Thomas Cleland of Cornell University and Nabil Imam of Intel in San Francisco set their neuromorphic AI against the traditional neural network in a 10-odour smell test. The neuromorphic system sniffed a single sample of each odour to train. Traditional AI has undergone hundreds of training sessions to learn each smell. During the test, each AI sniffed samples in which the learned smell was only 20 to 80 per cent of the overall scent — mimicking real-world conditions in which the target smells are often intermingled with other scents. Neuromorphic AI had the right smell 92 percent of the time. The standard AI has achieved 52 percent accuracy.

Priyadarshini Panda, a neuromorphic engineer at Yale University, is impressed by the neuromorphic AI’s strong sense of smell in muddled samples. The new AI’s one-and-one learning strategy is also more energy-efficient than traditional AI systems, which “tend to be very power-hungry,” she says (SN: 9/26/18).

Another advantage of the neuromorphic setup is that the AI can continue to learn new smells after its initial training if new neurons are added to the network, similar to the way that new cells continue to form in the brain.

As new neurons are added to the AI, new scents can be tuned without disrupting the other neurons. It’s a different storey for traditional AI, where the neural connexions involved in the recognition of a certain odour, or a set of odours, are more widely distributed across the network. Adding a new smell to the mix is likely to disrupt those existing connexions, so the typical AI struggles to learn new scents without forgetting others — unless it is re-trained from scratch, using both the original and the new scent samples.

To demonstrate this, Cleland and Imam have trained their neuromorphic AI and standard AI to specialise in toluene recognition, which is used to make paints and fingernail polish. The researchers then tried to teach the neural networks to recognise acetone, an ingredient of the nail polish remover. The neuromorphic AI simply added acetone to its fragrance-recognition repertoire, but the standard AI could not learn acetone without forgetting the smell of toluene. These types of memory lapses are a major limitation of the current AI (SN: 5/14/19).

Continual learning seems to work well with the neuromorphic system when there are few scents involved, Panda says. “But what if you’re going to make it large-scale? “In the future, researchers will be able to test whether this neuromorphic system can learn a much wider range of scents. “This is a good start,” she says.

Comments are closed.