- ... Schmidhuber1
- Current address:
Institut für Informatik,
Technische Universität München,
Arcisstr. 21, 8000 München 2, Germany
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...Prelinger:92
2
-
Simply maximizing the variance
of the output units without obeying constraint (6) will not
necessarily maximize the number of different classifications.
Example:
Consider a set of four different four-dimensional
input patterns 1000, 0100, 0010, 0001. Suppose the classifier
maps the first two input patterns to the four-dimensional
output pattern 1100 and the other two to 0011. This will
yield a variance of 4.
A `more discriminative' response would
map each pattern to itself, but this will yield a lower variance of 3.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
- ...
dependencies3
- Steve Nowlan has described an alternative
non-predictor based approach for finding
non-redundant codes [Nowlan, 1988].
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.