Purchase this article with an account.
Jonathan Victor, Ferenc Mechler, Ifije Ohiorhenuan, Anita Schmid, Keith Purpura; Neurons in primary visual cortex show dramatic changes in filtering properties when high-order correlations are present. Journal of Vision 2009;9(8):744. doi: 10.1167/9.8.744.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
V1 is widely considered to act primarily as a feedforward bank of filters followed by simple nonlinearities (“LN” systems). However, when LN models are built from simple analytically-convenient stimuli, their predictions of neural responses to natural scenes are only modestly accurate. One possibility is that this inaccuracy is merely quantitative, and can be remedied by adding gain controls, modulatory feedback, and multiple subunits to the basic LN structure. Alternatively, there might be fundamental qualitative differences between the computations performed by real cortical neurons and those performed by these models.
Since natural scenes have characteristics that traditional analytic stimuli lack, differences between responses of real and model V1 neurons likely reflect sensitivity to the distinguishing characteristics of natural scenes, namely, high-order correlations (HOCs). To isolate the effects of HOCs, we created sets of binary checkerboard stimuli in which second-order correlations were absent, but selected HOCs were present. Moreover, because our stimuli had identical contrast and spatial frequency content, they would equally engage cortical gain controls. For each of these statistical “contexts”, we determined the receptive field map - i.e., the L stage of the LN model that best accounts for the neuron's responses. Because stimuli were constructed so that second-order correlations were absent, these maps could be obtained by reverse correlation.
Recordings were made via tetrodes in four locations in macaque V1. In most (13/16) neurons, there were dramatic effects of high-order “context” on receptive field structure, including heightened sensitivity, development of spatial antagonism, or changes in response time course. In a few neurons, RF maps could only be obtained in the presence of high-order structure. These behaviors are not present for model LN neurons. This suggests that alternative model structures, such as strongly recurrent networks, are required to account for V1 responses even at a qualitative level.
This PDF is available to Subscribers Only