Definite Feature

A definite feature is a characteristic that is shared by all patterns that fall into the same band of hidden unit activity, where this band has been revealed as being distinct using a jittered density plot. Definite features are discovered by identifying all the input patterns that belong to the same band, and then performing a statistical analysis of the input features.

A definite unary feature is revealed by computing means and standard deviations of input features. If the standard deviation is 0, then all patterns in the band have a constant value -- a constant value equal to the computed mean.

A definite binary feature is a relational feature that is revealed by computing correlations amongst input features. A perfect correlation of 1 means that two different features are always equal for patterns that belong to the band; a perfect correlation of -1 means that the two different features are always not equal (assuming binary representation of inputs).

The discovery of definite features has been shown to be instrumental in interpreting the internal structure of multilayered networks of value units (Berkeley et al., 1995; Dawson et al. 2000).

References:

1. 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.
2. 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.