Sorry for so many questions… I cannot promise if this is my last one… .
So I have a Nengo network of K
neurons which requires N
input values and outputs one value each timestep. The N
inputs are supposed to be the outputs of N
neurons. I want to put this network between two Ensembles, but it comes with a catch. Just like the MaxPool
op, each of the N
inputs to this network from the previous Ensemble should be grid shaped and the single outputs should also be arranged in a grid (such that when flattened, it can be inputted to the next Ensemble or can have direct connection to the next Ensemble neurons).
Therefore, taking the pictorial example above, the 4 x 4
square represents the Ensemble neurons arranged in grid, there will be 4
instances of my Nengo network (each instance having its own K
neurons) and each of the N
neurons from the coloured sub-grid will be connected to an instance, with the 4
outputs forming the grid which will be directly input to the next Ensemble of 4 neurons (one-to-one connection).
The desired architecture can be seen in the picture below.
All the blue coloured
Nengo Nets
are independently functional and receive input in grid format, and their output is directly connected to the next Ensemble. Of course each of the inputs to the blue Nengo Nets
are neuron’s input and their (i.e. Nengo Nets
) output is fed to the individual neurons in next 2x2
Ensemble.
Keeping in mind the number of filters in a Conv Ensemble (previous and next both), how do I code this architecture? The reason I want such an architecture is to allow the neurons in the Nengo Nets
to maintain their (voltage) state in each timestep of the simulation.