In studies of visual perception performance is often measured by statistics that are ratios of a perceptual magnitude to its intrinsic variability, most commonly the signal-detection measure d'=delta sigma. Many models for visual phenomena treat the variability sigma as a constant and describe performance exclusively by delta. However, in models for the combination of stimulus attributes, the combination process affects both terms, and an observed d' reflects both delta and sigma. For example, we have shown that masking and configural effects with sinusoidal plaids can be at least partially interpreted as noise effects. We have developed methods to analyse these effects. Through a series of concurrent-response experiments using grating stimuli, some reported at earlier ECVP meetings, we have measured the form and magnitude of the noise sources. Our analysis allows us to model the way that primitive Fourier components (spatial frequency by orientation) are integrated to form second-order or third-order combinations (eg spatial frequency pooled over orientation).
An Optimum RBF Network for Signal Detection in Non-Gaussian Noise
