Go to the previous, next section, table of contents.
A natural gradient ica rule that can be bound to the underlying network. The rule is defined in terms of a set of pre-synaptic nodes, a set of post-synaptic nodes, a set of connections between the pre- and post-synaptic nodes. Optionally, there can also be a set of bias nodes and bias connections to the post-synaptic nodes. For cnset objects to be valid, the last record must contain only a set post-synaptic nodes with no connections. All other records must contain a single pre-synaptic node and a corresponding set of outgoing synapses. The plastic parameter is given by <parameter>. If <bias> is true, the the last cnset object defines the bias nodes in the system. Priors are defined using objects of type Distrib. A default prior can be defined both for the whole structure and for individual post-synaptic nodes. The <wfactor> parameter defines the relative weighting between the momentum term and the backward summation term.
Naming
prior/<n> references the nth priorConfiguration
wfactor=<data> sets the w_ij factor (1) prior=<ref> sets the default prior prior:<i>=<r> sets the prior for post-synaptic node iSource Points
yhat:<i> yhat value <i> uvalue:<i> u value <i> backprop:<i> backprop value <i>Additional arguments to existing commands
bind [parameter=<id>] [bias=<bool>] prior=<ref> <further arguments for derived commands> binds itself to the underlying network structure parameter=<id> the synaptic parameter which is plastic bias=<bool> whether the last cnset object defines a set of bias nodes prior=<ref> defines the object that is the default prior
Go to the previous, next section, table of contents. Document generated Mon Jun 16 02:19:05 GMT 2003