Then we asked: does this principle also apply to voices?
Using a recent dataset with >600 voice clips, we tested whether wav2vec embeddings showed the same effect.
👉 They didn’t. No consistent link between L2 norm and voice memorability.

And crucially:
This effect held even after controlling for word frequency, valence, and size.
So representational magnitude is not just a proxy for familiar or emotionally loaded words.

Then we asked: is this just a visual trick, or is it present in other domains as well?
When we turned to words, the result was striking:
Across 3 big datasets, words with higher vector magnitude in embeddings were consistently more memorable, revealing the same L2 norm principle

We first wanted to examine whether we could replicate this L2 norm effect as reported by Jaegle et al. (2019).
Using the massive THINGS dataset (>26k images, 13k participants), we replicated that the L2 norm of CNN representations predicts image memorability.