The precursor of this work (SRA 1.0) was published earlier this year on JML, titled "Informativity enhances memory robustness against interference in sentence comprehension":
www.sciencedirect.com/science/arti...

Our account naturally indicates that high-surprisal inputs require more memory resources. This suggests that the tendency to put more effort into processing surprising inputs may reflect an evolved strategy of our cognitive system that is adapted to an efficiency problem.

We also provide mathematical derivation (in Appendix) showing that the optimal encoding precisions to minimize total retrieval error should satisfy that high surprisal inputs be encoded with higher precision compared to low surprisal ones.

High-surprisal inputs are less likely to be reconstructed under the influence of prior, resulting in higher retrieval error. Therefore, to minimize total retrieval error across multiple inputs, more surprising ones need to be encoded with higher precision.

Second, since the true state of a past input is inaccessible, memory retrieval, or decoding, is effectively an inferential process that reconstructs a past input from uncertainty, using the statistical structure of long-term knowledge.

First, memory representation is noisy, and the true state of a past input is inaccessible. Importantly, following previous studies in psychophysics, we posit that more memory resources in encoding lead to less representational uncertainty of the input.

We propose that more memory resources should be allocated to more surprising linguistic units as the principled solution to a computational problem of working memory: how to maximize memory accuracy under the constraint of limited resources?

Fine-tuning a ggplot's aesthetic sensibilities is a time-honoured way to procrastinate from doing something you'd hate even more