design.txt

With the separation of sensory input and motor output into separate structures
in order to prevent them from directly interfering with one another, it raises
some complications with regards to connectivity and learning. Whereas previously
there could exist a single synaptic matrix which encoded the weights between
all neurons in the network, we must now come up with a new mechanism that will
allow for the network to continue to learn.

First, why did we choose to separate input/output from the main network?
The simple answer is that STDP was causing motor neurons to form connections
which stimulated input neurons -- this is obviously not what we want, and it is
not something which typically happens (to my knowledge) in biological nervous
systems. It would be very unfortunate (and impractical) if attempting to move
your left toe caused you to experience smell, for example.

Now, in order to allow the network to determine what neurons should receive
sensory input, and what neurons should be connected to motor output, I propose
the following:

In addition to the existing synaptic matrix for the recurrent network that is
intended to drive behaviour, there will also exist two additional matrices,
which will also be trained through STDP; one of which will be used for sensory
input, and one for motor output. In theory, this should prevent backwards flow
from motor output back to sensory input. An example of the matrices are as
follows:

Input:                  Recurrent:         Output:
|ir1 ir2 ir3|           |r1 r2 r3|      |ro1 ro2|
|ir4 ir5 ir6|           |r4 r5 r6|      |ro3 ro4|
                        |r7 r8 r9|      |ro5 ro6|

Each column in the input matrix corresponds to an input neuron; every entry in a
patricular column represents that input neurons synaptic connection to a neuron
in the recurrent network (of which there are 3 in this example). Likewise, for
every neuron in the recurrent network there is a column, which a synaptic weight
for every neuron in the output layer