Auryn simulator

Simulator for spiking neural networks with synaptic plasticity

User Tools

Site Tools


Complex Matrix class

As of Auryn v0.8 the library comes with a complex matrix class which generalizes SimpleMatrix to synaptic connections which can have more than just a single scalar value per connection. ComplexMatrix is essentially a rank three tensor class optimized for sparse synaptic transmission.

More specifically, like SimpleMatrix the complex matrix class implements a sparse matrix that allows fast access of all elements of one row. It provides row iterators to efficiently propagate spikes. Memory has to be reserved when the class is defined and elements can only be inserted row by row starting from “top left to bottom right”. This scheme enables related data fields to reside in memory next to each other.

ComplexMatrix generalizes SimpleMatrix to a rank 3 tensor in which each synaptic connection can have more than one value (third tensor mode – corresponding to a synaptic state). This allows to efficiently implement state based complex synaptic models which have their own internal dynamics. Instead of storing all synaptic data values in one long array the synaptic state values are stored in multiple state vectors which can be manipulated efficiently in a vector based manner.

For instance, suppose w holds your instance of ComplexMatris in which you have set_num_synapse_states(3). This makes it a connectivity matrix which reserves three state variables of type AurynWeight per synaptic connection. Now, decaying all elements of, say, state 3 by a factor foo is as simple as w→get_state_vector(2)→scale(foo). Adding two of the states for all synapses becomes


As you imagine these expressions can quickly become rather lengthy. It is therefore nice to declare shortcuts in your plastic Connection class such as:

AurynSynStateVector * w_val        = w->get_state_vector(0); 
AurynSynStateVector * tagging_val  = w->get_state_vector(1); 
AurynSynStateVector * scaffold_val = w->get_state_vector(2);

You can now work with these as you are used with AurynVector or AurynStateVector instances w_val→saxpy(foo,tagging_val)

manual/complexmatrix.txt · Last modified: 2016/06/16 19:43 by zenke