Fig. 4: Analyses of representational change.

A Schematic of representational similarity matrices (RSMs) derived from the Similarity-Based Word Arrangement Task (SWAT) procedure before learning (purple RSM) and after learning (orange RSM); note that our real matrices would be 60×60 words rather than the 5 × 5 words used in this toy example. Using the pair GENDER-FEMALE for illustrative purposes, we illustrate four of our key analyses. A. Analyses of pairwise representational changes across learning. In Analysis 1, cells outlined in yellow highlight the pairwise distance of the cue word GENDER to its target FEMALE, and we compare how this distance changes across learning. In Analysis 2, we examine the change in pairwise distance across learning between cue words (e.g., GENDER) and semantically related non-target words (lures; green outlines). B Analyses of individual word representations across learning. Pink outline reflects cue word GENDER, blue outline reflects target word FEMALE. We define the representation of an individual word as its row vector from the RSM (i.e. by its pairwise relationships to all other words in our set). In Analysis 3, we test how the representation of each word changes across learning by taking the Pearson correlation of the row vectors from the pre- and post-learning RSMs. In Analysis 4, we test whether the word representations in the to-be-learned pair change asymmetrically. In this analysis, we correlate the representation of the cue word before learning with that of the target word after learning and the representation of the cue word after learning with that of the target word before learning. The difference between these two values is calculated as a measure of asymmetry, where a positive value reflects the target being drawn towards the cue, a negative value reflects the cue being drawn towards the target and a value of zero reflects the cue and target being drawn towards each other symmetrically (or no representational change).