Neuroscientists at New York University and the Salk Institute have developed a new technique for measuring visual responses to complex images. The method consists of building a model based on cell responses to a range of stimuli, then asking how accurate the model is by comparing the model predictions with the actual responses of a cell to additional stimuli. Their work is reported in the most recent issue of the journal, Neuron.
The eye’s retina represents the visual world point-by-point-much like pixels on a computer screen render images. The cerebral cortex plays a fundamental role in this process by converting the "pixellated" representation into meaningful image components, such as lines, edges, and motions. Historically, researchers have sought to map out this process by measuring how visual cortical neurons-nerve cells that make up the brain’s cortex-respond to specific stimuli. However, researchers capture the processing of only a small fraction of visual reality under this method because it examines neurological responses to selected stimuli rather than how neurons function more broadly.
In an effort to overcome these limitations, Nicole Rust, J. Anthony Movshon, and Eero Simoncelli of NYU’s Center for Neural Science, as well as Odelia Schwartz of California’s Salk Institute, used randomly generated images to more fully explore all possible stimulus combinations. The researchers used novel statistical techniques to determine what aspects of the random images caused an increase or a decrease in the activity of a neuron. By studying the origins of visual processing as a whole, rather than tracing how cortical neurons respond to selected stimuli, the analysis uncovered a set of previously unsuspected mechanisms that work in concert to a range of cortical responses.
To accomplish this, Rust and colleagues employed a relatively new technique, the analysis of spike-triggered covariance, which is used to measure the activity of neurons, in order to predict cortical responses much more accurately than classical models. By building precise mathematical models that describe exactly how the response of a neuron depends on the stimulus presented to it, they were able to demonstrate that these neurons are more complex than previously appreciated.
"This analysis can be extended to study the further transformations of image components into the higher representations that allow us to recognize, remember, and act upon visual information," said Rust.
New York University. June 2005.