Representational magnitude as a geometric signature ofimage and word memorability
What makes some stimuli more memorable than others? Recent work has shown that image memorability is predicted by the magnitude of population responses in both monkey inferotemporal cortex and convolutional neural networks, suggesting that stimuli that more strongly activate distributed feature representations are more likely to be remembered. However, it remains unclear whether this effect is specific to the visual domain or reflects a more general property of distributed representations. Here, we show that representational magnitude predicts memorability beyond the visual domain: the effect not only replicates for images in an independent dataset, but extends to lexical memory. Across three large-scale word datasets, word embedding magnitude reliably predicts recognition memorability, independent of word frequency, valence, or length. However, the effect did not extend to a recently published voice memorability dataset, possibly reflecting distinct features driving auditory memory. Together, the cross-domain generalization suggests that the representational magnitude effect is a general property of distributed representations. Memorability, on this view, is inherent to encoding: stimuli that activate more features, and activate them more strongly - whether in brains or artificial neural networks - leave a larger representational footprint, and therefore a more lasting memory trace.