Olfaction Through the Lens of Representational Alignment

Abstract

Human olfactory perception emerges from complex interactions between the chemical properties of odorants, neural processing, and individual experiences. Understanding how these different levels of information relate to one another remains a fundamental challenge. Recent advances in machine learning have introduced models that learn representations from large amounts of chemical and linguistic data, raising the question of whether these representations capture information relevant to human olfaction without explicit perceptual supervision. In this thesis, we use representational alignment as a common framework to investigate how molecular and linguistic representations relate to human perceptual measurements and neural activity. Across three studies, molecular transformers and large language models are evaluated against different forms of human olfactory measurements, including categorical labels, continuous ratings, similarity judgments, and functional magnetic resonance imaging (fMRI) responses. These findings demonstrate that pretrained molecular representations contain information relevant to both human perceptual judgments and neural responses. Large language models also show substantial alignment with human olfactory judgments, sometimes outperforming molecular models, although their relative performance depends on the task and input representation. Neural analyses further highlight differences in the olfactory information captured by pretrained representations. Through a complementary study, we examine the robustness of deep learning methods for electroencephalography (EEG) analysis under data corruption, demonstrating that improvements in signal reconstruction do not necessarily translate into better downstream classification. Together, these studies demonstrate that general-purpose pretrained models capture information relevant to multiple aspects of human olfaction, while highlighting how measured alignment depends on the representations, response types, and evaluation criteria. This thesis contributes to understanding the relationships between chemical, linguistic, perceptual, and neural representations, providing a foundation for further investigations of human olfactory processing using pretrained models.

Link to DiVA