Abstract
Perceptual learning changes the efficiency of observers' internal calculations (1,2). However, the exact nature of those changes has remained a mystery. Here, we use response classification (3) to measure observer calculations directly as learning takes place. The result is a set of ‘classification images’ that reveal how the linear components of an observer's calculations change with learning. Observers identified unfamiliar low contrast versions of either two male faces or two abstract textures embedded in high contrast Gaussian white noise. For each stimulus type, there were 12 sessions of 800 trials per observer, conducted over the course of 2-3 weeks. Contrast thresholds were measured during each session with an adaptive staircase procedure, which also served to maintain the signal-to-noise ratio throughout the experiment. Thresholds improved by up to a factor of 5 across the first 6 sessions and changed very little during the last 6 sessions. The correlation between human and ideal classification images increased by up to a factor of 4 between the first and the last half of the experiment, reflecting the increase in calculation efficiency with learning. In addition, the number pixels that reached statistical significance (p<0.01) in the human classification images increased by as much as a factor of 2, indicating that at least part of the changes that took place with learning involved increases in the area over which information was used. These effects produced visible shifts in both the size and distribution of features within the classification images, and are consistent with previous findings of an improvement in calculation efficiency with perceptual learning.
1.
GoldJBennettPJSekulerAB Nature 1999, 402:176–178.
2.
DosherBALuZL Vision Research 1999, 39:3197–3221.
3.
AhumadaAJLovellJ JASA 1971, 49:1751–1756.