Manifold learning for olfactory habituation to strongly fluctuating backgrounds

François X. P. Bourassa, Paul François, Gautam Reddy, Massimo VergassolabioRxiv 2025 · DOI

TL;DR

This paper proposes a "manifold learning" approach for olfactory habituation, suggesting that the brain learns to subtract fluctuating background odors by projecting them onto a low-dimensional subspace. This mechanism, implemented by biologically plausible neural networks, outperforms traditional predictive filtering methods in noisy, turbulent environments and offers a new perspective on how sensory systems adapt to complex natural scenes.

What Problem Does This Paper Address?

Animals rely heavily on their sense of smell, but natural environments present a significant challenge: important odor signals are often masked by a cacophony of background odors that fluctuate wildly due to atmospheric turbulence. This makes it difficult for the olfactory system to distinguish a novel, behaviorally relevant odor from the ever-changing olfactory "noise." While some adaptation mechanisms are known, how the system effectively filters out these dynamic backgrounds to detect new scents remains unclear. This paper tackles this challenge by proposing a novel computational framework for olfactory habituation that accounts for the high-dimensional nature of odor coding and the specific statistical properties of turbulent odor environments.

Methods

The authors developed a computational model of the early olfactory pathway, inspired by the structure of the insect antennal lobe and mammalian olfactory bulb. The core of their model is a network of inhibitory interneurons that learn to represent the "manifold" (a low-dimensional subspace) spanned by background odors. They explored two biologically plausible plasticity rules for these interneurons:

  • IBCM (Intrator, Bienenstock, Cooper, Munro): A plasticity rule inspired by visual cortex learning, which allows neurons to become selective for specific input components.
  • BioPCA (Biologically Plausible Online Principal Component Analysis): A rule that learns the principal components of the input.

These rules are used to train the inhibitory interneurons to project incoming odor signals onto the learned background manifold and subtract this projection, effectively removing the background. The model's performance was evaluated by its ability to recognize a new odor after habituation to a fluctuating background. They also derived analytical solutions for optimal predictive filtering and manifold learning strategies to establish a theoretical baseline for comparison.

Datasets

The paper does not use traditional "datasets" in the machine learning sense. Instead, it simulates olfactory environments with specific statistical properties:

  • Turbulent odor statistics: Mimicking natural atmospheric conditions with intermittent "whiffs" and "blanks" of odor, characterized by power-law distributions of durations and concentrations.
  • Weakly non-Gaussian and Log-normal backgrounds: Used for specific analytical and simulation tests to explore the robustness of the proposed mechanisms.

These simulated environments serve as the "data" for training and evaluating the model's habituation and odor recognition capabilities.

Key Insights and Ideas

  • Manifold Learning as a Solution: The paper argues that the high-dimensional nature of olfactory coding and the low-dimensional structure of background odor mixtures naturally lend themselves to a "manifold learning" approach for habituation. The brain learns the "shape" of the background odor space and subtracts signals that lie within it.
  • Beyond Mean Subtraction: Traditional methods like subtracting a time-averaged background are insufficient for highly fluctuating, turbulent odors. Manifold learning offers a more robust solution.
  • Inhibitory Interneurons as Key Players: The model highlights the crucial role of inhibitory interneurons in the early olfactory pathway for implementing this manifold learning strategy.
  • IBCM and BioPCA as Plausible Mechanisms: Both the IBCM and BioPCA plasticity rules, when applied to inhibitory interneurons, effectively learn the background manifold and lead to significant improvements in odor recognition.
  • Specificity for Background Components: IBCM neurons, in particular, show a tendency to become selective for individual background odor components, effectively learning a basis for the background space.

Key Results

  • Superior Performance: Manifold learning models (IBCM and BioPCA) significantly outperformed average subtraction and no-habituation baselines in recognizing new odors amidst turbulent backgrounds.
  • Robustness: The proposed mechanisms were shown to be effective across a range of olfactory space dimensions and odor concentrations, and robust to OSN noise.
  • Analytical Support: Theoretical analysis confirmed that manifold learning is optimal in regimes with fast background fluctuations and high-dimensional odor spaces, characteristic of olfaction.
  • Biological Plausibility: The IBCM and BioPCA rules are grounded in known biological plasticity mechanisms, making the proposed solution biologically relevant.

Limitations and Future Work

  • Linear Manifolds: The current model primarily assumes linear manifolds for background odors. Future work could explore more complex, curved manifolds, which might require more sophisticated algorithms like manifold tiling.
  • Simplified Network Architecture: The model focuses on early olfactory processing and omits neuromodulatory inputs and higher-level cognitive functions like attention, which are known to influence olfactory perception.
  • Experimental Validation: The authors suggest that experimental tests could involve imaging PN activity to observe habituation-induced decreases in norm and variance, and observing how interneuron activity tracks background stimuli.
  • Broader Applications: The principles of manifold learning for habituation could be applicable to other sensory systems and even to figure-ground segregation in time-varying signals, such as in video object detection.

Summary updated

Abstract

Animals rely on their sense of smell to survive, but important olfactory cues are mixed with confounding background odors that fluctuate due to atmospheric turbulence. It is unclear how the olfactory system habituates to such stochastic backgrounds to detect behaviorally important odors. Here, we explicitly consider the high-dimensional nature of odor coding, the natural statistics of odor fluctuations and the architecture of the early olfactory pathway. We show that their combination favors a manifold learning mechanism for olfactory habituation over alternatives based on predictive filtering. Manifold learning is implemented in our model by a biologically plausible network of inhibitory interneurons in the early olfactory pathway. We demonstrate that plasticity rules based on IBCM or online PCA are effective at implementing this mechanism in turbulent conditions and outperform previous models relying on mean background subtraction. Interneurons with an IBCM plasticity rule acquire selectivity to independently varying odors. This manifold learning mechanism offers a path towards distinguishing plasticity rules in experiments and could be leveraged by other biological circuits facing fluctuating environments.
BibTeX
@misc{bourassa2025manifold,
  author = {François X. P. Bourassa and Paul François and Gautam Reddy and Massimo Vergassola},
  title = {{Manifold learning for olfactory habituation to strongly fluctuating backgrounds}},
  year = {2025},
  doi = {10.1101/2025.05.26.656161},
  howpublished = {bioRxiv},
}
Figure