Abstract
Previous studies show systematic biases when visual features are remembered over short delays. However, some studies find biases toward ensemble statistics such as the mean of all presented features (attraction bias), while others find the opposite (repulsion bias). Here we investigated the factors that determine the direction of bias and whether biases have a functional benefit. When individual item representations are less reliable, ensemble-level representations may be more useful. However, when individual items are more reliable, repulsion biases could help reduce confusion between similar representations. To evaluate repulsion or attraction biases, we used a continuous color report task with two memory items. We independently manipulated the fidelity of the color targets (encoding times of 50/150/500ms) and the distance between the targets in feature space (differences of 20°/45°/90°/135°). Bias direction was quantified as the proportion of reports away from the un-probed item. When encoding time was relatively long (500ms), repulsion became monotonically stronger as color distance decreased. This suggests that subjects exaggerated the distance between increasingly similar items when they were higher fidelity. At a shorter encoding time (150ms), the repulsion biases instead peaked when the targets were 45° apart. This suggests that the fidelity of two very similar colors (i.e. 20° apart) became too low to register as distinct representations. Lastly, when encoding time was short (50ms), repulsion biases deceased overall – with representations attracted toward the average hue. This indicates that when fidelity is low, people relied on group-level representations. We provide evidence that the degree of repulsion vs. attraction depends on the strength and the similarity of the individual representations. This is consistent with the idea that such biases are adaptive ways for the visual system to overcome its biological limitations, allowing systematic errors to occur to maximize the usefulness of concurrent representations.
Meeting abstract presented at VSS 2018