Representational magnitude as a geometric signature of image and word memorability

What makes some stimuli more memorable than others? While memory varies across individuals, research shows that some items are intrinsically more memorable, a property quantifiable as “memorability”. Recently, a new representational signature for image memorability was identified: the magnitude of the population response in convolutional neural networks (CNNs) correlated with image memorability. However, it is unclear if this geometric principle was confined to the visual domain or whether it represents a broader computational phenomenon observable in other stimuli domains as well. Here we show that this representational magnitude effect not only replicates for images in an independent dataset but also generalizes to an entirely different cognitive domain and neural network architecture: lexical memorability and word embeddings. Across three large-scale lexical datasets, we found that the L2 norm (vector magnitude) of word embeddings reliably predicted recognition memorability, independent of word frequency, valence, or word length. This consistency suggests the effect reflects a general property of distributed representations, where representational magnitude may capture how strongly a stimulus projects onto dominant, conceptual/semantically meaningful features in a network’s embedding space. At the same time, this effect does not appear to generalize to all domains, as our analysis of a novel voice memorability dataset showed no such relationship between representational magnitude and memorability. Together, these findings indicate that the geometry of distributed representations offers a useful lens for understanding memorability, suggesting an item’s representational magnitude reflects its projection onto dominant dimensions of representation, in both artificial and biological systems.