- The research found two types of neurons: “reliable cells” that identify distinct odours and “unreliable cells” that help distinguish similar scents with experience.
- The variability in neural response was found to come from a deeper circuit in the brain, suggesting it serves a significant purpose.
- This neural variability might benefit continual learning systems in AI, making them more discerning.
How can wine experts pick out different citrus or tropical fruit aromas or those of different flowers?
Cold Spring Harbor Laboratory (CSHL) associate professor Saket Navlakha and Salk Institute researcher Shyam Srinivasan may have the answer. They have found that certain neurons allow fruit flies and mice to tell apart distinct smells.
The team also observed that with experience, another group of neurons helps the animals distinguish between very similar odours.
The study was inspired by research from former CSHL assistant professor Glenn Turner. Years ago, Turner noticed something odd. When exposed to the same scent, some fruit fly neurons fired consistently while others varied from trial to trial.
At the time, many researchers dismissed these differences as a product of background noise. But Navlakha and Srinivasan wondered whether the variations might serve a purpose.
“There were two things we were interested in,” Navlakha says. “Where is this variability coming from? And is it good for anything?”
To address these questions, the team created a fruit fly smell model. The model showed that the variability came from a deeper circuit of the brain than previously thought. This suggested the variation was indeed meaningful.
'Unreliable cells are useful'Next, the team observed that some neurons respond differently to two very dissimilar odours, but the same to similar smells. The researchers called these neurons reliable cells. This small group of cells helps flies quickly distinguish between differing odours.
Another much larger group of neurons responds unpredictably when exposed to similar smells. These neurons, which the researchers call unreliable cells, might help us learn to identify specific scents in a glass of wine, for example.
“The model we developed shows these unreliable cells are useful,” Srinivasan says. “But it requires many learning bouts to take advantage of them.”
Of course, this research isn’t just for wine drinkers. Srinivasan says the results might help explain how we learn to differentiate between similarities detected by other senses, and how we make decisions based on those sensory inputs.
The findings could also lead to better machine-learning models. Unlike fruit fly and mouse neurons, computers generally respond the same to the same inputs.
“Maybe you don’t want a machine-learning model to represent the same input the same way every time,” Navlakha explains. “In more continual learning systems, variability could be useful.”
That means this research could someday help make AI more discerning and reliable.
This article was originally published on the CSHL Newsstand.