Abstract
Theories of feature-based attention postulate gain modulation of sensory neurons that encode target features to enhance processing of task-relevant information (e.g., Treue and Martinez-Trujillo, 1999). More recently, Navalpakkam and Itti (2007) demonstrated that target representations are also shifted to optimize the distinctiveness of targets to distractors (see also Becker et al., 2010). For example, orange targets may be represented as being more "reddish" if distractors are predictably "yellowish". However, it remains unclear whether shifting target representations away from distractor features is the only way to reduce distractor interference. We examined this question in 3 experiments (each N=40) using a visual search task. A target defined by a single color was presented with a single distractor that differed from the target-color by 0 to 60 degrees in L*a*b color space (5 degree increments). The probability of "high-similarity" vs. "low-similarity" distractors was manipulated between-groups by determining their presentation frequency from a mirror-reversed half-normal distribution. The group with a greater probability of seeing high-similarity distractors had shorter RTs and higher accuracy, but only for highly similar distractors. Importantly, in a simultaneous target identification task, both groups showed a similar shift in representing the target hue as being more distant from the distractor set (i.e., counter-clockwise from the target color). However, subjects in the high-similarity group were also less likely to identify as targets colors that were used as highly similar distractors. This suggests that being exposed to more similar distractors produced asymmetrically "sharper" target tuning. In sum, these experiments suggest that the target representation is always shifted to optimize distinctiveness from the distractor set, but that increased exposure to highly similar distractors further sharpens the target representation to more effectively suppress highly similar distractors.
Meeting abstract presented at VSS 2016