Feature repulsion depends on task demands

Next we asked wether the repulsion of similar color values depends on the task demand to separate the objects or if it is a byproduct of any associative learning. In the previous experiments, the associative memory task required participants to distinguish between the two similar colored objects. Here, we ran experiment that was identical to Experiment 1, expect we changed the task during learning so that optimal performance required integrating across the overlapping face-object pairs (Figure 2.4A). Specifically, we introduced an inference test in which one of the learned faces appeared and participants had to select the face (from a set of 4 faces) that was paired with the same object. With this subtle task change, discriminating the color of the objects was no longer relevant to performance on the task. Rather, optimal performance required linking between the two faces in memory via their shared association with the common object (irrespective of its color). We then tested how this change in task demand - from separating the colors to generalizing across them - would effect the repulsion bias. Would the bias remain, diminish, or possibly flip resulting in an attraction bias reflecting a merging of the color values? 
Participants successfully learned to infer the face pairings as their selection of target face increased over learning rounds (F(1,25) = 225.3, p< 0.00000001; Figure 2.4B) and by the end of learning they were above chance performance  in all conditions (ps <0.000001). The similarity of the colors did not impact generalization across the object-face pairs as average inference performance across rounds did not vary by similarity condition (F(2,50) = 0.52, p = 0.67). Despite not being explicitly tested on their ability to discriminate between the competing colors, subjects' color error decreased over learning  (F(1,25) = 101, p <0.0000001; Figure 2.4C).   
After 8 rounds of learning color error varied by similarity condition (F(2,50) = 6.34, p = 0.003). Follow up t-tests revealed that there were larger color errors for objects in the low similarity condition compared to the both the high similarity (t(25) = -2.90, p = 0.008) and the moderate similarity conditions (t(25) = -2.55, p=0.017). However, color error did not differ between the high and moderate similarity conditions (t(25) = 1.48, p=0.15). Turning to our main question of interest we asked if the change in task demand altered the bias in color estimates. As in Experiment 1 color bias varied by similarity condition (F(2,50 = 5.18, p = 0.009), however, in this experiment we found no color bias in the high similarity condition (t(25) = -0.48, p = 0.63). Surprisingly, both the moderate similarity and high similarity  conditions  similarity condition showed positive bias towards the competing color(moderate: t(25) = -2.09, p = 0.047; low:  t(25) = -5.15, p = 0.00003) . Directly comparing the bias measures across Experiment 1 and Experiment 3 revealed a significant difference in color bias for  the high similarity  (t(47.42) = 3.52, p = 0.001) and low similarity conditions (t(46.70) = 3.54, p = 0.0009) conditions but not in the moderate similarity condition (t(47.31) = 1.61, p = 0.11). Thus, changing the task demand to encourage integration across pairs eliminated the repulsion bias in the high similarity condition and induced an attraction bias in the  low similarity condition. 

Discussion

Summary
While there have been numerous studies demonstrating that learning reduces interference between similar memories, to date, there have been no studies examining if and how the features of those corresponding memories change over learning. Using a novel behavioral paradigm that assessed feature memory on a continuous scale, we show that learning induced a repulsion bias in color memory between competing memories such that the colors were remembered as being more different from each other than their true colors. Importantly, this repulsion bias developed over learning and was competition dependent as it was only observed for colors that were highly similar (24 degrees apart). Furthermore, the repulsion of competing colors was adaptive to memory performance as greater repulsion between the colors was associated with reduced interference between them. These findings provide insight into how memory representations get distorted over learning to reduce interference between them.  
Our study design in that objects were paired with we were not interested in just wether participants could remember the pairings but how the remembered them. 
This research contributes to our knowledge of how competitive retrieval shapes memories. Competiion based frameworks of memory retrieval propose that retrieving a target memory employs inhibitory mechanisms that acts to suppress any competing traces that might interfere with successful recollection (Anderson 2003, Levy 2008). However, this inhibition has critical and lasting consequences on the non-target memories it suppresses. Behavioral research has demonstrated that the act of retrieval of a target memory makes it less likely for the competing, non-target memories to be subsequently remembered (Anderson RIF). A recent neuroimaging study confirmed that multiple retrieval attempts of a target memory suppressed the cortical representation of competing memories (Wimber). Intriguingly, in paradigms where participants receive an opportunity to restudy the previous competing items, memory for those items improves relative to items that were studied the same number of times but had not been the competing item during retrieval (Storm 2008, Hulbert 2014). Thus, it has been shown that repeated testing and restudy opportunities of similar events can quantitatively facilitate their future retrieval. The current set of studies adds to this research by demonstrating how competitive learning can qualitatively change those memories. Here, we found that learning exaggerated the features (color) of  competing memories apart and furthermore that this exaggeration reduces the interference between them. This suggests that the inhibition processes engaged during competitive retrieval might not fully suppress the entire competing memory trace, but rather selectively inhibit the shared features of the memories (Norman papers) . 
The repulsion bias observed for the features of memory, strongly parallels recent fMRI findings that initial overlap in hippocampal patterns triggers of repulsion of their patterns with learning. Prior studies have found that the hippocampus can come to exaggerate the differences between similar memories by coding them more distinctly from each other than to unrelated memories. A key component in all these studies is that this effect emerged after extensive learning and experience with the stimuli. Indeed, studies that directly measured the change in hippocampal patterns with learning found that similar memories initially had overlapping activity patterns within the hippocampus but after learning those patterns were repelled away from each other (Chanales). This accrual of repelled representations over learning parallels the emergence of repelled color memories in this study. Whereas initial color estimates tended to be biased towards the competitor, after learning color estimates were biased away from the competing color. Furthermore, fMRI studies have demonstrated that repulsion occurs when competition between memories is maximal. Similarly, in this study, the repulsion bias in color memory occurred in the high similarity condition in which the competing colors were very similar.  when the initial colors were very similar. and the fMRI Future work will need to test if the two results are related. 
An intriguing result from these studies is that the repulsion bias depends on the task demand to separate. When the task in between study rounds was changed to encourage integration across the face-object pairs, the repulsion bias in the high similarity condition disappeared. This suggests that top-down signals interact with the mnemonic processes engaged during learning to shape the organization of information in memory storage. This corroborates research suggesting that memory representations flexibly store the features most relevant to current goals (Mack, Love). Thus distancing the features of memory should only be adaptive in as much as discrimination between the memories is beneficial for future behavior.  Although we demonstrated that the demand to separate is critical to mechanistically repel the features of memories, it is unclear at what stage of processing this demand contributed to the repulsion. It could be the goal to separate altered the attentional weighting afforded to the diagnostic feature relevant for separation (i.e. the color of the objects) which in turn enhanced its encoding (Aly, Mack). Or possibly the task demand influenced downstream processes known to influence the degree of separation such as the strength of reactivation of the competing memory (Norman, Kim) or relatedly ,the level of inhibition employed during retrieval (Anderson). Future work will need to further investigate how the top down goals dictated by the task demands contribute to the  repulsion effect. 
An important question this research raises 
Lastly, this research contributes to a growing body of research demonstrating that repulsion of similar information is a common coding scheme employed by the mind and brain. While this is the first evidence of a repulsion bias in the featural content of long term memory, similar biases have been found at many levels of processing from perceptual representations in visual working memory (Lae, Rademaker) to estimates of temporal duration (Ezzyat, Huesser), to judgements of social categories (Krueger, Forster, Wilder). This confluence of findings across a wide range of fields strongly suggests that distortions exaggerating the differences between similar representations is not an error arising from computational or resource limitations but rather is a feature of the system that improves discrimination performance between otherwise confusable information. Future work will need to explore whether repulsion biases across these different domains arise from a common mechanism or whether each domain developed a distinct mechanism to arrive at a shared repelled representational structure. 

Methods

Experiment 1

Exclusions