Global spatial representations anchor local data to stimulus configuration

Local visual representations seem to account for the effects of target location on recall precision. However, we also found that interactions among the elements that make a visual scene affect recall accuracy. These effects do not seem to be spatially limited, but rather increase with their distance from the recalled target (Hubbard & Ruppel, 2000). When the observers had to remember only one item, the recalled target location was displaced eccentrically, that is away from the centre of the display. However, when three items had to be remembered, the recalled target location was displaced toward, rather than away from centre of the display. The direction change of the recall bias with memory load was found to reflect two linearly independent distortions. The first was a hypermetric bias in display coordinates, which solely determined the nature of the systematic error when one item was memorised, and a hypometric bias toward the centre of the memory items’ configuration, which dominated the systematic error when three items were memorised. However, the persistence of the hypermetric bias in screen coordinates when three items were held in memory, suggests separate causes for the hypermetric and hypometric biases. Hence, we suggest that spatial memory depends on representations which employ separate reference frames, one in display coordinates and the other relative to the centre of the set of memorised elements. Other interpretations of the hypometric bias are not particularly plausible. For example, it is possible that participants may have occasionally reported the location of a non-target memory item, giving the impression that the recalled location of the target was displaced, when averaged over trials, toward the centre of the memory items’ configuration. However, we excluded from the analysis trials in which the

recalled location was closer to one of the non-target memory items than the target. Moreover, when the recall display contained two of the memory items, the target was still recalled toward the centre of the memory items’ configuration, albeit by a diminished amount, indicating that when the proportion of binding errors was minimised, analytically or by the probe procedure used to identify the target, the hypometric bias in memory coordinates persisted. However, the most conclusive evidence was provided by the finding that translating the non-target items at recall caused a parallel translation of the recalled target location by a similar amount, a finding that simply cannot be explained by participants confounding target with non-target items. A second hypothesis one may entertain, is that the location of the memory items is coded in relative terms, namely that observers keep track of the distance between items in memory rather than their location in relation to some other reference. If this were the case, then any rigid displacement of the non-target memory items should result in displacement of the recalled target to preserve the overall configuration. However, when we displaced the non-target memory items at recall, by rotating their position around an axis passing through the centre of the display and orthogonal to the image plane, the recalled target location was not displaced in the direction that would have preserved the distance between the target and non-target items, but rather in the opposite direction. This finding fundamentally undermines the view that observers were memorising the distance between the target and non-target memory items, and supports instead the interpretation that the target position is coded and recalled relative to a common reference frame, namely the centre of the memory items’ configuration. Thirdly, it has been suggested that configurational effects on recall may be best understood within a Bayesian framework (Brady & Alvarez, 2011; Brady, Konkle, & Alvarez, 2011). Accordingly, the configuration of the memory array is used to generate a prior for the distribution of the possible target positions, which, combined with noisy estimates of the actual target position, leads to target recalls biased toward the average position of the memory items. This proposal would predict an increase in centripetal bias

whenever information about the memory configuration is provided. However, we found that presenting the non-target items at recall diminishes the centripetal bias. Single unit data in higher order Parietal areas of behaving primates have indicated that in a match to sample task, visually evoked responses to previously seen dot stimuli are largely invariant to parallel displacements of the stimulus over the retina, suggesting a stimulus centred coding of complex spatial configurations (Chafee et al., 2005; Chafee, Averbeck, & Crowe, 2007), such as those that may underpin the behavioural effects summarized above. The maintenance of the configuration of memory arrays in higher order regions could also account for the finding that the complexity of the memory array scales the amplitude of delay period activity in Parietal and Frontal regions during vWM tasks (Todd & Marois, 2005; Vogel & Machizawa, 2004).

In conclusion, both behavioural and neurophysiological data are consistent with the idea that there are at least two representations of spatial data in vWM - one where fine spatial details are maintained by spatially curtailed mechanisms in early Visual cortex, the other where summaries of the global configuration are maintained by mechanisms that generalise over space in higher cortical areas.