Abstract
The discovery of repeated patterns (e.g., AABB in the string AABBCBCAABBCAABB) enables us to form compressed representations that reduce processing demands. We have developed a novel paradigm for investigating how the human visual system discovers spatial patterns (specifically, translated point collections). Stimuli consisted of approximately 35 gray points arranged on a 16x16 (hidden) grid (see PDF, Fig. 1). The display always contained 4 embedded point collections on 4x4 grids, some of which were translations. In a 6s discover phase, subjects identified point collections that occurred translated elsewhere in the display. The difficulty of identifying translations was manipulated by (1) adding random points or "noise" (low and high noise conditions with signal-to-noise ratio of 10 and 2 respectively) and (2) varying the distance between embeddings (low and high separation conditions with minimum 1 and 4 grid distance). Immediately following, subjects saw a point collection in a 4s decide phase, and responded whether or not that collection had occurred more than once in the previous discover phase (Figs. 2-3). A within-subject ANOVA on accuracy, with factors for noise and separation, revealed a significant main effect of noise (F1,5 = 24.2, p <.01, see also Fig. 4). Eyetracking data suggested that in the low noise condition, subjects perceived the 4 embedded collections on a global scale first (Fig. 1, fixations 1-3). Next they checked for differences between pairs of embeddings on a local scale (Fig. 1, fixations 7-11), with tracks similar to visual comparison tasks (Pomplun et al., 2001; Galpin & Underwood, 2005). Often, global perception was impeded by high noise, leading to incorrect responses. An adapted pattern discovery algorithm (Meredith et al., 2002) had higher accuracy on this task than all but one subject. This is likely because the algorithm, which does not employ a global-local strategy, is less affected by noise.
Meeting abstract presented at VSS 2013