Neural networks set sights on visual processing in brain.

Neural networks set sights on visual processing in brain

The past few years have witnessed an explosion in scientific attempts to develop computer models that simulate the behavior of small groups of brain cells involved in functions such as vision and smell (SN: 1/24/87, p.60). According to a report in the Feb. 24 NATURE, one such "neural network" has demonstrated the ability to code visual information much in the way that has been observed among monkey brain cells concerned with estimating the position of visible objects. David Zipser of the University of California at San Diego and Richard A. Andersen of the Massachusetts Institute of Technology in Cambridge say that the cortex, or outer layer of the brain, and their computer model may handle incoming information similarly.

"This is one of the first applications of neural network technology to experimental data from the brain," says biophysicist Terrence J. Sejnowski of Johns Hopkins University in Baltimore, who has designed a similar computer simulation of neuron activity in the cat brain. "We can apply this type of model to different parts of the cortex as an all-purpose tool for studying brain function."

Zipser and Andersen, as well as Sejnowski, use a neural network training procedure called "back propagation." The system contains a layer of input units, a layer of output units and an intermediate or "hidden" layer of units, which, with repeated trials, takes on response properties that best accomplish the computational task being learned. As training proceeds, error signals are sent back through the network to adjust the strengths of connections between all units in order to nudge the system toward a desired output.

In this instance, the trained responses of the hidden layer were compared to electrical measurements taken from a small area of monkey cortex containing neurons that track the visual field and eye position. Lesion studies indicate that monkey neurons in this region combine information about the position of an object on the retina of the eye with information about the direction in which the eyes are pointing; this helps determine the object's location in relation in the body.

The model network was trained using randomly selected pairs of input eye positions and retinal positions. The true spatial location implied by each pair of inputs was programmed into the model and generated error signals that produced accurate spatial estimates within about 1,000 trials.

In learning to carry out this task, say the researchers, the system modified itself so that "hiden unit" responses to visual input closely matched electrical responses of critical monkey neurons when the animals view an object. This supports the notion, they add, that the brain carries out a number of steps in determining where an object is, including the combining of retinal and eye position information.

Sejnowski's neural network uses back propagation training to compute curvature from shading in an image, an important part of depth perception. Input units in the network are arranged to mimic the activity of visual receptor cells in the cat. After training, hidden units acquire properties much like those of cells in the cat's visual cortex that are sensitive to elongated shapes. In addition, output units behave like another class of neurons that further process information about shapes.

"These little networks are not models of the brain per se," says Sejnowski. "But we can develop networks that help to understand the functioning of particular circuits in the brain. Just as calculus can be applied to problems in a variety of disciplines, a back propagation network can be applied to the study of different parts of the cortex."