Abstract
When given a pattern detection task, an analysis of the subjects' internal templates or classification images reveals a significant difference between signal present and signal absent conditions. Six combinations of two cosine gratings, A and B (1 c/deg and 3 c/deg) were used as the target stimulus patterns (A, B, A+B, A−B, A−(1/3)B and A+(3/4)B). The one-dimensional noise was the sum of 1st–7th harmonics of a 0.5/c/deg fundamental. We used the method of constant stimuli with 4 test contrasts (first level was 0 or no signal) and 4 responses. The stimulus and noise were designed to allow a reconstruction of the subject's internal template, the weighting of sensitivity to each frequency component, analyzed with linear regression. Because subjects don't know if the signal will be present or not, they should always be looking for the same target pattern. This predicts that the internal template should be nearly identical regardless of signal presence or absence. Results from five subjects show that this is clearly not the case; in the signal absent conditions, the template efficiency (correlation of ideal template and human template) drops significantly (from ∼70% efficiency to ∼30%). However, the correlation of the ratings using a double pass method was found to be approximately 0.45, independent of the stimulus level. There also appears to be an effect of frequency, subjects were more efficient in detecting the B grating in isolation than any of the other patterns. Our classification image findings imply that when there is not enough signal to reinforce a more accurate template, the subject is unable to use their internal template correctly, possibly confusing similar frequency patterns with the target. The surprising implication from our double-pass experiments is that these mismatches or mistakes are done consistently.