For each sub-band, we create a confusion matrix (CM), which is simply
an extended matrix of *hits* and *misses* for all classes. To
summarize the confusion matrix, we calculate mutual information (MI)
for each CM [8] as
,
where *i* is the feature we would like to
transmit, and *j* is the feature that is perceived. We estimate the
probabilities *p*_{ij}, *p*_{i}, and *p*_{j} from *n*_{ij}/*n*, *n*_{i}/*n*, and
*n*_{j}/*n*, respectively, where *n*_{i} is the frequency of stimulus *i*,
*n*_{j} is the frequency of response *j*, and *n*_{ij} is the frequency
of the joint occurrence of stimulus *i* and response *j* in a sample
of *n* observations.