An integration device is a type of processing unit that can be used in an artificial neural network. Named according to Ballard's (1986) terminology, an integration device uses a monotonic activation function that increases activity as net input increases. This activation function is usually sigmoid-shaped, and therefore is sometimes called a "squashing" function. Mathematically, it is usually defined using the logistic equation, and ranges between 0 and 1. However, other versions are possible.

Integration devices are the most common component in the artificial neural networks that are used in cognitive science. A multilayered network of such devices is usually trained with the generalized delta rule discovered by Rumelhart, Hinton, and Williams (1986). Introductions to these devices, and discussions of their capabilities and limitations, have been provided by Dawson (2004, 2005).

References:

1. Ballard, D. (1986). Cortical structures and parallel processing: Structure and function. The Behavioral And Brain Sciences, 9, 67-120.
2. Dawson, M. R. W. (2004). Minds And Machines : Connectionism And Psychological Modeling. Malden, MA: Blackwell Pub.
3. Dawson, M. R. W. (2005). Connectionism : A Hands-on Approach Malden, MA: Blackwell Pub.
4. Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning representations by back-propagating errors. Nature, 323, 533-536.