Abstract
Detection and discrimination of signals in external noise can provide a wealth of information about suprathreshold processing. Recently a number of researchers (including us) have used external noise to calculate the classification image that is presumably an estimate of the observer's template. Here we call into question the fixed template framework.
We measured the detection and discrimination of a Gabor-like patch in the presence of external noise using a double-pass method. Our grating stimulus was: cos(y)^10 cos(12y) for y going from −0.5 to +0.5 deg. This is the sum of 11 harmonics from 1 to 11 c/deg. The superimposed noise was the sum of the same 11 harmonics with random amplitudes and random phase. The rms noise contrast was equal across components and varied over a wide range across runs. A rating scale method of constant stimulus was used. Linear regression of the observer's responses onto the eleven cosine noise components gives an accurate classification image (with far fewer trials than with the cross-correlation method). The experiment is repeated using identical random noise and identical stimuli. The consistency of the responses to the identical stimuli is used to estimate the ratio, R, specifying the consistent vs. the random component of multiplicative noise. An alternative method for calculating R is based on the ratio of the observer's d′ divided by the d′ of an ideal observer using the observer's classification template.
Our experimental results show that for both detection and discrimination, R, determined by the double pass method is about half of R determined by the d' ratio. That is, the double pass method shows much more consistency of responses than would be expected from the observers' d′. The simplest explanation is that rather than there being a single template, the multichannel model is still operating at suprathreshold levels. Different channels are optimal on different trials producing a consistent, but inefficient response.