Why do we remember some stimuli better than others? Previous research has shown that some words and images are consistently more memorable than others, suggesting that memorability is partly related to properties of the stimuli themselves. More recent research suggests that memorability may also be related to how stimuli are represented within a broader representational space. This study investigated whether representational features can predict memorability, and whether this relationship generalizes across computational models, encoding tasks and sensory modalities.

First, word memorability was predicted using semantic representations from GloVe to examine whether a relationship previously found with Word2Vec also generalizes to another embedding model. Second, the role of encoding task was examined by comparing deep semantic encoding with shallow non-semantic encoding. Third, the same approach was extended to the visual domain using high-level visual representations from a pretrained ResNet-50 model. Across analyses, representational features were found to contain information relevant to memorability. This relationship generalized across semantic embedding models and was also observed in the visual domain. However, the predictive value of these representations depended on the encoding task, with semantic representations being more informative under conditions that required semantic processing.

Together, these findings demonstrate that representational features contain information relevant to memorability across computational models and sensory modalities, while their predictive value varies depending on the encoding task.