Purchase this article with an account.
James H. Elder, Yaniv Morgenstern; Power spectrum classification image analysis reveals localized mechanisms underlying nonlinear detection of narrowband stimuli. Journal of Vision 2006;6(6):117. doi: 10.1167/6.6.117.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Prior measurements of detection thresholds for gratings in noise have suggested that detection may be based upon linear cross-correlation with relatively small broadband filters (Kersten, 1984). Recent direct measurements of spatial summation using a classification image technique have yielded contrary results, showing broad spatial summation over many cycles of the narrowband stimulus (Morgenstern & Elder, 2005). Here we report a computational model that partially resolves this contradiction. We show that under the standard linear cross-correlator model, classification images derived from signal-present trials yield broad summation fields locked to signal frequency and phase. However, classification images derived from signal-absent trials reveal no peak at signal frequency, indicating that the linear detection model does not apply (Ahumada & Beard, 1999). We test a number of alternative models, and show that an energy model based on broad spatial pooling of responses from local broadband mechanisms is most consistent with the human data. We show that this model is linear in the power spectrum domain, and introduce a novel classification image analysis technique that allows direct estimation of the two-dimensional bandpass transfer function of the underlying local mechanisms. These mechanisms are found to be highly localized in space, quantitatively similar in their spatial and spatial-frequency properties to neurons in early visual cortex of primate, and qualitatively similar to the broadband mechanisms inferred indirectly by Kersten (1984).
This PDF is available to Subscribers Only