Package pylearn :: Package shared :: Package layers :: Module rust2005
[hide private]

Module rust2005

source code

Provides Rust2005 layer

Paper: 

This layer implements a model of simple and complex cell firing rate responses.


:TODO: implement full model with variable exponents.  The current implementation fixes internal
exponents to 2 and the external exponent to 1/2.

:TODO: add the weights on the quadratic filter
responses.  E and S are supposed to be the "square root of a *weighted* sum of squares".
The simplifications here are probably useful, but make them optional, or a different class or something.
- The current implementation can be interpreted as including the weights inside the filters.
  The filters are not constrained to have a unit norm, for example.

Classes [hide private]
Rust2005
Energy-like complex cell activation function described in Rust et al.
Rust2005Conv
Convolutional version of `Rust2005`
Functions [hide private]
 
rust2005_act_from_filters(linpart, E_quad, S_quad, eps)
Return rust2005 activation from linear filter responses, as well as E and S terms
source code

Imports: numpy, Image, theano, shared, softsign, softplus, ConvOp, update_locals, add_logging


Function Details [hide private]

rust2005_act_from_filters(linpart, E_quad, S_quad, eps)

source code 

Return rust2005 activation from linear filter responses, as well as E and S terms

:param linpart: a single tensor of linear filter responses :param E_quad: a list of tensors of linear filter responses :param S_quad: a list of tensors of linear filter responses :param eps: a scalar to add to the sum of squares before the sqrt