Abstract
In everyday perception, we are surrounded by large numbers of objects with variable features. Instead of interpreting this variety as completely different objects, we can quickly categorize them as belonging to one or several types (e.g., berries and leaves). We suggested earlier that such categorization can be based on ensemble statistics: The visual system "decides" whether the overall feature distribution has one or more "categorical" peaks. We have tested this theory for unidimensional feature distributions (Utochkin & Yurevich, 2016). Here, we test it for conjunction-defined ensembles. In three experiments, observers discriminated between two textures filled with lines of various lengths and orientations. Length and orientation had same distributions in the textures, but their correlations were opposite. The crucial manipulation concerned the shapes of feature distributions: They could be "segmentable" (only extreme feature values presented with a large gap between them), or "non-segmentable" (both extreme and middle values presented with smooth transition between). In Experiment 1, we found that segmentable displays yield steeper psychometric functions indicating better discrimination. In Experiment 2, we found that the effect of segmentability on texture discrimination occurs at 100-200 ms and requires both feature dimensions having a "segmentable" distribution. We interpreted this in terms of "preattentive" division of the textures into categorical classes of conjunctions. That is, the visual system (1) divides the set into subgroups based on highly distinctive peaks, (2) selects all items from one of the peaks, and (3) decides whether a second dimension also forms two peaks across textural patches. In Experiment 3, we tested a hypothesis that imperfect selection of a subset at step (2) is an important limiting factor of accurate categorization. When a half of lines from one side of one of the feature distributions were removed, discrimination using the remaining feature distribution was much better.
Meeting abstract presented at VSS 2018