Abstract
Suppose you are searching satellite images of North Korea for launch sites, steel plants, bridges, etc. More mundanely, suppose you are searching for anyone you know at a party. How does search proceed when you are looking for any one of many possible targets held in memory? In work in the 60 s–70 s, observers typically searched through 1–4 visible letters for any of 1–4 remembered letters. Effects of visual and memory set size were each linear and inefficient so the combination of visual and memory search was inefficient X inefficient. But, surely, your search for any friend in a crowd does not involve a serial search through the visible faces convolved with serial search through the set of all friends held in memory. Something must change when the number of items in the relevant memory set substantially exceeds working memory capacity. In our experiment, ten observers memorized sets of 1–16 photorealistic objects. They then performed 300 searches for any of those objects in visual displays with set sizes 2–16. One target was always present. Observers identified targets with mouseclicks. The effects of visual set size were linear for all memory set sizes. The effects of memory set size, however, are decidedly non-linear. RT increases with the log of memory set. Four visual set sizes X five memory sets yield 20 RTs. These 20 are a linear function of visual set size X log2(memorySet) + 1) The average slope of this function is 23.5 msec. Average r-sq:0.94 (no log, r-sq = 0.77). If logarithmic search through large memory sets generalizes, these values would allow you to find one of 1000 friends in a crowd of 100 in about 30 seconds (ignoring such issues as eye movements and navigating the party). By contrast, linear searches through memory with the same 23 msec/step slope would take about 40 minutes.