Olfactory perception requires mapping unique chemical information to our internal representations of objects (e.g., strawberry), categories (e.g., food), and concepts (e.g., pleasantness). However, in contrast to other sensory modalities, it is not well understood which core dimensions form our olfactory space and how to reliably discover them. Our approach to address this question is based on the organizational scheme of representations in terms of (dis)similarities.
In a first step, each of our participants (N = 8) performed three tasks with 16 different odorants to derive such dissimilarities: (1) feature rating of each odorant according to 8 features, (2) rating of pairwise dissimilarities between odorants (120 pairs) and (3) the triplet-odd-one-out task (TOOO), where participants choose out of an odorant triplet the one that differed most from the others (560 triplets). The three experiments were repeated after at least one week’s break. For each task, we quantified pairwise dissimilarities from the responses to obtain representational dissimilarity matrices (RDMs). Crucially, the TOOO task revealed the most informative RDMs and the highest retest reliability out of all three tasks.
In a second step, we trained a concept embedding model on the TOOO data given that we may assess olfactory representations best with this task. The model provided interpretable and reproducible dimensions characterizing the odorants and predicted dissimilarity structures with high accuracies using only a subset of possible triplet combinations for training. Overall, our findings demonstrate the potential of AI to undersample in the TOOO and to discover the dimensionality of our olfactory space.