Perceptual metrics for odorants: Learning from non-expert similarity feedback using machine learning

Area of Science:

• Chemosensory science
• Computational chemistry
• Psychophysics

Background:

• Current odorant similarity metrics rely on physicochemical properties or discrete verbal labels, which inadequately capture the nuances of olfactory perception.
• Both molecular structure and simplistic descriptors fail to represent the continuous nature of odor perception.
• There is a need for more granular and data-driven approaches to quantify perceptual similarity in olfaction.

Purpose of the Study:

• To introduce a novel data-driven method for learning perceptual similarity metrics for odorants.
• To leverage user perceptual feedback alongside physicochemical features for improved odor dissimilarity modeling.
• To develop a more effective representation of the continuous odor perception space using relative comparisons.

Main Methods:

• Implemented perceptual metrics learning (PMeL) by combining odorant physicochemical features with user perceptual feedback.
• Utilized triplet comparisons (e.g., 'Does A smell more like B or C?') from users to gather relative similarity judgments.
• Trained models on non-expert feedback to evaluate perceptual dissimilarity between odorants.

Main Results:

• Demonstrated the effectiveness of the PMeL approach in evaluating perceptual dissimilarity across various tasks.
• Showcased that user-relative similarity comparisons provide a more granular representation of the continuous perception space.
• Validated the model's alignment with expert similarity judgments, reducing reliance on expert annotations.

Conclusions:

• The proposed data-driven PMeL method effectively quantifies odorant perceptual dissimilarity.
• User-based relative comparisons offer a scalable and effective way to model continuous odor perception.
• This approach enhances our understanding of olfactory mechanisms and reduces the need for extensive expert annotations.