Summary: Scientists have uncovered how animals distinguish closely related odors by combining stable and variable neural responses. This work reveals how some neurons reliably signal distinct smells while others introduce trial-to-trial variability that helps animals learn to tell very similar scents apart over time.
The study, based on experiments and computational modeling of olfactory circuits in fruit flies and mice, shows that neural variability is not merely noise but arises from deeper brain circuitry and supports experience-dependent discrimination. The findings point to new ideas for improving machine-learning approaches to continual learning and sensory recognition.
Key Facts:
- The research identifies two functional classes of neurons in olfactory circuits: “reliable cells” that produce consistent responses to different odors, and a larger population of “unreliable cells” whose responses vary across trials.
- Modeling and circuit analysis indicate that the observed variability originates from deeper layers of the olfactory network, suggesting an adaptive role rather than simple background noise.
- Incorporating controlled variability into artificial neural systems could improve continual learning and make AI better at discriminating highly similar sensory inputs.
Source: CSHL
How do humans and animals separate similar scents? When a sommelier describes wine as having citrus, floral, or tropical notes, the brain is performing a complex task: extracting and comparing subtle odor features. Yet, many scents overlap, and to the untrained nose they may all smell like “wine.” New research from Cold Spring Harbor Laboratory (CSHL) Associate Professor Saket Navlakha and Salk Institute researcher Shyam Srinivasan sheds light on how brains resolve these fine-grained olfactory differences.
Building on earlier observations by former CSHL Assistant Professor Glenn Turner, the team examined how individual neurons in fruit flies and mice respond when the same odor is presented repeatedly. Turner’s work had shown that some neurons fire consistently while others vary across repetitions. Instead of dismissing that variability as mere noise, Navlakha and Srinivasan investigated whether it plays a functional role in olfactory discrimination and learning.
To explore the origin and consequence of variable responses, the researchers developed a computational model of the fruit fly olfactory system and compared model predictions to neural recordings. The model pointed to deeper network layers as the source of response variability, implying that this variability is a structured feature of the circuit rather than random measurement error.
Results revealed two complementary neural strategies for odor discrimination. A relatively small set of “reliable cells” generates distinct, stable responses to clearly different odors, enabling rapid, high-confidence identification. In contrast, a much larger set of “unreliable cells” responds inconsistently to very similar odors, showing trial-to-trial variability that by itself is ambiguous but becomes informative after repeated exposures and learning.
Srinivasan emphasizes that unreliable cells are not dysfunctional; instead, their variable responses provide a substrate for gradual learning. “The model we developed shows these unreliable cells are useful,” he explains. “But it requires many learning bouts to take advantage of them.” In other words, unreliable responses help refine discrimination through experience, allowing animals to detect subtle differences that a single, consistent response pattern might miss.
Beyond basic olfaction, the authors suggest this principle could generalize to other sensory systems where distinguishing subtle differences improves behavior and decision-making. The interplay between stable signals that support immediate recognition and variable signals that support gradual learning likely contributes to flexible, experience-dependent perception.
The findings also have implications for artificial intelligence and machine learning. Most current computational models produce the same representation for the same input every time, which can limit adaptability in continual learning scenarios. Navlakha notes, “Maybe you don’t want a machine-learning model to represent the same input the same way every time. In more continual learning systems, variability could be useful.” Introducing controlled variability might help AI systems better separate highly similar inputs and learn from ongoing experience.
In sum, this research reframes neural variability from unwanted noise to a purposeful feature of olfactory circuits—one that supports nuanced sensory discrimination and learning. The study provides a mechanistic account rooted in experimental data and modeling, and it opens a pathway for designing more flexible, biologically inspired machine-learning systems.
About this olfaction and neuroscience research news
Author: Samuel Diamond
Source: CSHL
Contact: Samuel Diamond – CSHL
Image: The image is credited to Neuroscience News
Original Research: The findings will appear in PLOS Biology