[theano-users] Constraints on convolutional filters

[Sym]

I would like to add a constraint on the filters of my convolution operator,
such that they'll always be symmetric and positive semi definite along the
two trailing (spatial) axes.

There are many ways to achieve this, for instance taking the Gaussian
Kernels, or building the symmetric Gram Matrices of the filters.

To ensure that this conditions is always met throughout the network
training, I'm applying this transformation to the filters before every
convolution.

The filters will then be modified between an update and a forward step ! I
am not sure it's a good idea...



*My questions :*Will the network be able to learn if the filters are passed
through this transformation at each forward step ?

Is there another way to constrain the filters of a layer on some condition?