Value Unit

A value unit is a type of processing unit that can be used in an artificial neural network. Named according to Ballard's (1986) terminology, a value unit is tuned to respond to a narrow range of net inputs. Net inputs within this range produce high responses; the unit generates weak responses to net inputs that fall either below or above this range. In other words, a value unit behaves as if it has two thresholds: a low one and a high one. Dawson and Schopflocher (1992) use a version of the Gaussian equation, with a minimum of 0, a maximum of 1, and a standard deviation of 1, to define the activation function of a value unit.

The generalized delta rule was modified by Dawson and Schopflocher (1992) to permit multilayer networks of value units to be trained. They found that these networks often learn faster than do traditional networks (i.e. networks constructed from integration devices). It was also discovered that value units behave in such a way that their responses are "banded"; this property has been exploited to interpret the internal structure of networks of value units (Berkeley et al., 1995; Dawson, 2004; Dawson et al., 2000).

References:

1. Ballard, D. (1986). Cortical structures and parallel processing: Structure and function. The Behavioral And Brain Sciences, 9, 67-120.
2. Berkeley, I. S. N., Dawson, M. R. W., Medler, D. A., Schopflocher, D. P., & Hornsby, L. (1995). Density plots of hidden value unit activations reveal interpretable bands. Connection Science, 7, 167-186.
3. Dawson, M. R. W. (2004). Minds And Machines : Connectionism And Psychological Modeling. Malden, MA: Blackwell Pub.
4. Dawson, M. R. W., Medler, D. A., McCaughan, D. B., Willson, L., & Carbonaro, M. (2000). Using extra output learning to insert a symbolic theory into a connectionist network. Minds And Machines, 10, 171-201.
5. Dawson, M. R. W., & Schopflocher, D. P. (1992). Modifying the generalized delta rule to train networks of nonmonotonic processors for pattern classification. Connection Science, 4, 19-31.