Rich Caruana and I have collaborated to investigate a similar effect in standard supervised algorithms. In the back-propagation algorithm, new stimulus dimensions can be used as other inputs to expand the input space, or as extra constraints to be satisfied by the network mapping from the other inputs to the desired outputs. We have found in several real problems where high dimensional input patterns are presented, there are some dimensions of the input patterns that are very useful as inputs, and others that are best discarded as inputs but used instead as extra outputs (and others that are best discarded period). Future work will investigate the aspects of the statistical relationship between dimensions that determine how they should best be used. This work will provide computational hypotheses of how sensory modalities should interact to develop the best representations for invariant object recognition.