Abstract
Intro
Our memories are not veridical snapshots of the past, rather, our memories are stored adaptively in order to best guide future behavior. One example in which experiences are stored adaptively in memory is when two experiences are very similar and easily confusable. Research has shown that in these instances, the hippocampus serves to distort the representations of the similar experiences by separating them into distinct neural patterns in order to minimize the interference between them (REFS). Recently, a number of studies (Chanales 2017, Favila 2016, Hulbert 2014, Schlichting 2015, Kim 2017) have provided evidence that over extended learning the hippocampus can actually repel the representations of similar experiences apart to the point where two similar events are coded more distinctly than two unrelated events. In this way the hippocampal code exaggerates the differences between overlapping events. Although this hippocampal repulsion has been shown to benefit future learning by reducing the confusability of the corresponding memories (Favila 2016), it not known if this repulsion that occurs at the neural level also occurs at the featural level of the corresponding memories. One intriguing possibility is that competition between similar levels can also repel the feature values of those memories apart such that those memories will be remembered as being more different from each other than they actually were.
The theoretical and empirical research conducted on the repulsion of hippocampal patterns suggest some critical factors underlying the emergence of this effect. First, the divergence of hippocampal activity patterns is a direct response to competition between the overlap in the events' neural patterns. The hippocampus resolves this competition by distancing the competing representations apart from one another. Therefore, the extent of repulsion should scale with the amount of competition between the similar events at the start of learning. Second, the divergence of the hippocampal patterns occurs over learning. As similar events continue to compete, the hippocampus will continue to distance their representations apart. Thus, a complete repulsion of the patterns will only be observed after repeated experience with the similar events.
Here, across multiple behavioral studies, we tested the idea that repeated encoding of highly similar stimuli would yield a similar repulsion of feature values in memory. We designed a novel paradigm in which participants performed an associative learning task that required them to discriminate between pairs of objects that were identical except for their color values. Critically, we adjusted the similarity of the object pairs by varying the difference in hue angle between the paired objects. After learning we probed participants memory for the color of each object using a continuous scale (REF). Assessing memory for the objets in this manner allowed us to measure whether color estimates were biased towards or away from the competing object's color.
This design allowed to test several predictions of the repulsion account. First, if repulsion also occurs at the featural level of memories, we should observe systematic distortions in color memory such that color estimates for each object would be biased away from the competing object's color. Second, if the repulsion of the feature values of memory is a response to competition, then maximal repulsion in color memory should occur for objects with highly similar color values. Lastly, we predicted that these repulsion biases should be adaptive to memory performance by decreasing the confusability of between the similar objects. Thus the the degree of repulsion between objects should be associated with decreased interference between them during a memory test.
Results
Similarity between items leads to a repulsion of their feature values in memory
In Experiment 1 participants completed an associative learning task in which they studied 36 pairs of colored objects and faces. To create competition between the pairs, the 36 objects each contained a competing object that was identical to its competitor aside from its color value (Figure 2.1A). To modulate the level of competition, we varied the difference in hue angle of the color between the competing objects across three conditions of color similarity: high similarity (24o difference), moderate similarity (48o difference), and low similarity (72o difference; Figure 1A). The object-face pairs were learned over 8 rounds of study and test. During a study round each colored object was presented with its associated face. Importantly, the two competing objects were never presented simultaneously, thus discrimination between them was entirely memory based. After each study round participants completed two tests: (1) a color memory test during which participants reported the color of each object using a color wheel, and (2) an object-face associative memory test during which subjects selected the face that corresponded to each object. On each test trial, the subject was presented with an object and four face choices: the face paired with object ('target'), the face paired with the competing object ('competitor'), and two faces paired with different objects ('non-competitor'). Immediately after all learning rounds participants completed an additional color memory test (post-test) and then returned, after a 24h delay, for a final color memory test (Day 2 post-test) (Figure 2. 1B).
We first turned to the associative memory test data to test that subjects successfully learned to discriminate between the similar object colors. Subjects learned to choose the target face over the learning rounds (F(1,22) = 435.4, p<0.0000001; Figure 2.2A) such that by the last round of learning they reliably chose the target faces over the competitor faces in all similarity conditions (ps <0.0000001). However, confirming that our similarity modulation influenced the discrimination difficulty between the colors, target hit rates across learning varied by similarity condition (F(2,44) = 13.04, p =0.00003). Average target selection rates were lower in the high similarity condition compared to both the moderate similarity (t(22) = -3.98, p = 0.0006) and low similarity conditions (t(22) = -4.65, p = 0.0001). Target selection rates did not differ between the moderate and low similarity condition (t(22) = -0.72, p = 0.48). This confirmed that increasing the similarity between colors increased the competition (interference) between them, particularly for the high similarity pairs. Turning toward the color memory test, color error for the objects steadily decreased over learning (F(1,22) = 166.2, p <0.0000001; Figure 2.2B) such that by the last round of learning there was no difference in color error between similarity condition (F(2,44) = 1.22, p = 0.31). Thus participants were able to learn each object's associated color across all conditions.
Next we assessed if and how color memory for the objects was distorted over learning. Specifically, we were interested in wether there was a bias in color memory for each object relative to its competing color. We quantified this color bias by calculating the percentage of trials in which the color estimate fell away from the competing color. Any deviations in the measure from 50% would reflect a bias in color memory with values greater than 50% reflecting a repulsion away from the competing color and values below 50% reflecting an attraction towards the competing color. We calculated this bias measure separately for each round and condition to see how the biases in color memory changes with experience (Figure 2C). There were significant changes in the color bias measures over learning across all the conditions (F(1,22) = 20.34, p =0.0002). At the start of learning color estimates tended to biased towards the competing color while towards the end of learning they were biased away from the competing color. Thus, there were dynamic shifts in the internal color representations of the objects over learning. However, to better quantify the bias in these representations and how they differed by similarity condition we turned to the post-test which was designed to have more trials and therefore more statistical power to detect condition differences.