Flexible updating of dynamic knowledge structures

Abstract

Schemas are knowledge structures that allow us to make efficient judgments about the world without the cost of memorizing every detail of previous experiences. It has long been known that schemas can enhance long-term memory for related information. The usefulness of schemas, however, critically depends on their adaptability: how flexibly a schema can be updated according to changing environmental conditions. Prior consolidation of a schema supports new learning of schema-consistent information. Yet, the effect of consolidation on inconsistent information, and how schemas may be subsequently updated, are not well understood. It is difficult to track the dynamic updating of knowledge structures with traditional memory measures. Here, using a continuous-report paradigm, we were able to show that schematization increases incrementally with consolidation and that the strength with which schemas are initially established predicts schema-guided responding in a later test. Critically, schema updating in response to inconsistent information was more pronounced in a group which was given time to consolidate compared to a group that was not given time to consolidate. Importantly, the later group reverted back to the no longer relevant schema, indicating that systematic bias towards old information, rather than increased forgetting, underlies reduced memory for schema-inconsistent information.

Figure 1

(A) Experimental phases for both groups across the sessions of the experiment. The no-consolidation group learned two categories of stimuli in the IS phase on day 1 and learned new stimuli of the same two categories during NL1. This group learned two new categories (the two categories that were not learned on day 1, or ‘irrelevant’ categories) on day 2, and came back for the final memory test on day 3. The consolidation group also learned two categories during the IS period on day 1, but in contrast to the no-consolidation group subsequently learned two new categories (the two categories that were not learned in the IS phase, or ‘irrelevant’ categories). They were presented with new stimuli of the same two categories that were learned during IS only on day 2, and came back for a final memory test on day 4. In NL1 for the no-consolidation group and NL2 for the consolidation group one of the two categories learned changed their schema mean: the mean shifted by 90 degrees clockwise compared to what it was in the IS phase. This category is referred to as the inconsistent category. (B) Left: Circle segment associated with the animal schema. Most of the animals are presented within the shaded area (not shown during the experiment). Middle: Inside-quadrant animal trial; the stimulus is presented within the area shown on the left. Right: Outside-quadrant animal trial; the stimulus is not presented within the area specified on the left. Image of Oia Church is a cropped version of the image from wikipedia.org, licensed under the Creative Commons Attribution-ShareAlike 3.0 License (https://creativecommons.org/licenses/by-sa/3.0/). All other images (pig and elephant, backgrounds removed) were retrieved from pixabay.com, licensed under a Creative Commons CC0 License (https://creativecommons.org/publicdomain/zero/1.0/deed.en).

Figure 2

Mean absolute error in inside-quadrant (green) and outside-quadrant (red) trials across experimental phases in the consistent category condition, for the no-consolidation group (lighter shades, circles) and the consolidation group (darker shades, squares). Error bars reflect standard errors of the mean.

Figure 3

(A–D) Schematic description of the hypothesized model components used for both the no-consolidation group (magenta) and the consolidation group (cyan) in Model 4. (A) Standard mixture model. (Note that (A) by itself is equivalent to Model 3). (B) In Model 4 a number of trials are described by a von Mises distribution centred on the old schema mean (on average −90 degrees from the target value). It is predicted on some trials that participants will revert to the old schema mean instead of random guesses or responding with the target value. This effect is predicted to be stronger in the no-consolidation group. (C) Hypothesized resulting distribution for trials of the inconsistent category, as a combination of the standard mixture model (A) and an additional von Mises distribution centred on the old schema mean (B). (D) Predicted (left) and observed (right) error distribution in the inconsistent category for the consolidation group (cyan, CG) and the no-consolidation group (magenta, NCG) overlaid on each other (overlap plotted in blue). Note the increased number of responses around the −90 degree location for the no-consolidation compared to the consolidation group.

Figure 4

Relationship between initial schema strength (measured as the difference in absolute error between outside- and inside-quadrant trials in the Initial Study phase, for the consistent or inconsistent category, respectively) and the probability of responding consistent with the ‘old schema’ in the Final Test for trials of the inconsistent NL category (A) and the probability of responding consistent with the relevant schema in the Final Test for trials of the consistent category (B). Left: Correlations for the no-consolidation group. Right: Correlations for the consolidation group.

Figure 5

(A) Example stimuli presented in the study phases of the experiment (here for animals and clothing). Stimuli were presented for 4 s, followed by 500 ms of a fixation cross (not shown). While most stimuli of a category fit a schematic location (inside-quadrant trials), some pictures are presented outside of this category-specific circle segment (outside-quadrant trials, here underlaid in red). (B) Illustration of the continuous response task. The stimulus appeared at a random location on the screen and the participant had 6 s to move the stimulus on the circle (here indicated by arrows which were not present in the actual experiment) to the remembered location using the left and right arrow key. Each trial was separated by a fixation cross presented for 500 ms (not shown). In the FT phase objects from each of the four categories were tested intermixed. In the test phases after the study phases in each block of IS and NL1/NL2 only the stimuli (and thus two categories) learned in the immediately preceding block were tested. Image of Oia Church is a cropped version of the image from wikipedia.org, licensed under the Creative Commons Attribution-ShareAlike 3.0 License (https://creativecommons.org/licenses/by-sa/3.0/). All other images (hat, boot, frog, jacket, pig, elephant, backgrounds removed) were retrieved from pixabay.com, licensed under a Creative Commons CC0 License (https://creativecommons.org/publicdomain/zero/1.0/deed.en).

Figure 6

Illustration of the schema shift in the inconsistent category. Shaded quadrants illustrate category-schema specific areas of the circle. (A) In this example the inside-quadrant trials of the ‘clothing’ category were presented in the left quadrant of the circle during the IS phase. (B) During the New learning phase (NL1 for the no-consolidation group, NL2 for the consolidation group) the category shifted by 90 degrees clockwise so that now inside-quadrant trials in the ‘clothing’ category would be presented in the top quadrant of the circle.