... entities1
The G-Max algorithm (Pearlmutter and Hinton, 1986) aims at a related objective: It tries to discover features that account for input redundancy. G-Max, however, is designed for single output units only.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... minimize2
Cross-entropy is another objective function for achieving the same goal. In the experiments, however, the conventional mean squared error based function (1) led to satisfactory results.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... error3
One might think of using Lagrangian multipliers (instead of arbitrary $\alpha, \beta, \gamma$) to rigidly enforce constraints such as independence. However, in order to use them the constraints would have to be simultaneously satisfiable. Except for special input distributions this seems to be unlikely (see also section 4.7).
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.


Back to Independent Component Analysis page.