Olfactory Signal Processing

Kush R. Varshney, Lav R. Varshney

Abstract

Olfaction, the sense of smell, has received scant attention from a signal processing perspective in comparison to audition and vision. In this paper, we develop a signal processing paradigm for olfactory signals based on new scientific discoveries including the psychophysics concept of olfactory white. We describe a framework for predicting the perception of odorant compounds from their physicochemical features and use the prediction as a foundation for several downstream processing tasks. We detail formulations for odor cancellation and food steganography, and provide real-world empirical examples for the two tasks. We also discuss adaptive filtering and other olfactory signal processing tasks at a high level.

Introduction

While signal processing (SP) for auditory and visual domains is highly mature, the sense of smell has remained a frontier for engineering. Unlike sound or light, which are characterized by waves with well-defined frequencies and amplitudes, olfaction involves complex chemical mixtures. This research, authored by Kush R. Varshney and Lav R. Varshney, establishes a formal mathematical framework to treat odors as signals, enabling operations like filtering and "cancellation" analogous to noise-canceling headphones.

Physicochemical-to-Perceptual Mapping

The foundational challenge in olfactory SP is bridging the gap between chemical composition and human perception. The authors propose a mapping where a compound’s physical properties (extracted from databases like PubChem) are represented as a vector $x$ and its perceived smell (across 146 descriptors such as "floral" or "sweet") as a vector $y$.

The relationship is modeled as a linear operator: $$y \approx Ax$$

To solve for the matrix $A$, the authors employ nuclear norm regularization. This approach is motivated by the observation that human odor perception is relatively low-dimensional—often dominated by a primary "pleasantness" axis. By enforcing a low-rank structure on $A$, the model avoids overfitting and captures the essential dimensions of olfactory perception while maintaining a manageable computational complexity.

Olfactory White and Active Cancellation

In acoustics, active noise cancellation involves adding a signal's inverse to achieve silence. In olfaction, molecules cannot be "subtracted" from the air. Instead, the authors utilize the concept of olfactory white. Much like white noise in audio or white light in optics, olfactory white is a neutral smell produced by mixing approximately 30 or more diverse odorants at equal intensities.

The authors formulate an optimization problem to identify the minimum number of compounds required to transform a specific malodor (e.g., durian, onion, or sauerkraut) into this neutral state. By adding specific "cleansing" chemicals, the resulting mixture is perceived as olfactory white rather than the original unpleasant scent. To ensure practical application, the paper uses group lasso (simultaneous sparsity) to identify a small, shared library of chemicals that can cancel multiple different odors.

Food Steganography

This paradigm extends to "food steganography," where the goal is to hide the sensory signature of an "averse" food (like broccoli) within a "delectable" cover medium (like mac and cheese). By adding a "steganographic key"—a specific mixture of ingredients—the model shifts the perceptual signature of the averse food toward a neutral point, effectively masking its presence without removing the nutritional content.

Empirical Results

The study synthesized data from the Atlas of Odor Character Profiles (AOCP) and the Volatile Compounds in Food (VCF) database to test the framework.

  • Predictive Accuracy: Five-fold cross-validation demonstrated a stable regularization parameter ($\lambda = 10^4$), which minimized error in the perceptual mapping and confirmed the validity of the low-rank assumption.
  • Odor Cancellation: Computational simulations showed that four distinct malodors (durian, onion, sauerkraut, and katsuobushi) could be effectively neutralized using a shared library of just 22 to 38 natural compounds.
  • Steganography Example: In a real-world simulation, the model identified Angelica Seed Oil as a primary ingredient for a food additive that could perceptually mask the "sickening/sulfidic" notes of cooked broccoli.

Discussion and Future Directions

The current framework relies on a linearity assumption for odor mixtures. While humans are largely "synthetic" smellers—perceiving complex mixtures as single objects rather than a list of ingredients—chemical interactions can be non-linear. Furthermore, unlike sound, smells linger and move through space. Future work must incorporate temporal dynamics, such as "Gaussian puff" models, to account for how odors persist over time. While these computational results are promising, the critical next step is human validation to confirm that the predicted "cancellation" matches subjective experience.

Summary updated

Abstract

Olfaction, the sense of smell, has received scant attention from a signal processing perspective in comparison to audition and vision. In this paper, we develop a signal processing paradigm for olfactory signals based on new scientific discoveries including the psychophysics concept of olfactory white. We describe a framework for predicting the perception of odorant compounds from their physicochemical features and use the prediction as a foundation for several downstream processing tasks. We detail formulations for odor cancellation and food steganography, and provide real-world empirical examples for the two tasks. We also discuss adaptive filtering and other olfactory signal processing tasks at a high level.
BibTeX
@misc{varshney2014olfactory,
  author = {Kush R. Varshney and Lav R. Varshney},
  title = {{Olfactory Signal Processing}},
  year = {2014},
  eprint = {1410.4865},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1410.4865},
}
Figure