In a modern multilayered perceptron, the output units do not have direct access to input signals, so they cannot be described as carving a pattern space into regions. Instead, they divide an alternate space, the hidden unit space, into decision regions (Dawson, 2004). The hidden unit space is similar to the pattern space, with the exception that the coordinates of the points that are placed within it are provided by hidden unit activities. That is, each pattern in a training set can be represented as a point in hidden unit space. The dimensionality of that space is defined by the number of hidden units in the network, and each hidden unit activity produced by an input pattern is one of the pattern’s coordinates.
- Dawson, M. R. W. (2004). Minds And Machines: Connectionism And Psychological Modeling. Malden, MA: Blackwell Pub.