Purchase this article with an account.
Nuno Goncalves, Andrew Welchman; Optimized computation of binocular disparity by populations of simple and complex cells. Journal of Vision 2017;17(10):756. doi: 10.1167/17.10.756.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Binocular disparities convey unique information about the depth structure of the environment. The process of capturing disparity by V1 neurons is typically described using the disparity energy model (Ohzawa et al., 1990). While this is convenient and elegant in its simplicity, the canonical model does not account for a number of neurophysiological observations. Principally it is not adapted to the statistics of natural disparities, and it fails to exhibit attenuation for binocularly anticorrelated stimuli. Here we develop a model system to test the optimization processes that might underlie neuronal encoding schemes. We used an unbiased modelling approach, optimizing a neural network to extract binocular disparity from naturalistic images. We then compared the properties of model's units with neurophysiological data. We found that a simple feed-forward neural network reproduces key aspects of V1 binocular neurons. First, simple units develop receptive fields with hybrid position and phase disparities, resembling the structure of V1 simple cells. We show that this scheme is adopted as it maximizes the Shannon Information about the depth of the scene. Second, complex units respond selectively to disparity in random-dot stereograms (RDS), without being trained on them. This selectivity extended to anticorrelated RDS, mimicking, very closely, the inverted and attenuated tuning curves found in many V1 complex cells. Using a stimulus optimization method, we show that complex units are maximally activated by position disparities, despite receiving activity from units that exploit phase encoding. Finally, we show that the encoding and readout mechanisms can be captured in simple analytical form, to produce model estimates that approximate the likelihood function for binocular disparity. These results provide a mechanistic and interpretive account of disparity processing in primates, and make a number of predictions for perceptual performance.
Meeting abstract presented at VSS 2017
This PDF is available to Subscribers Only