dwatola@DNS1.EECS.WSU.EDU (David Watola) (09/17/90)
i am trying to implement a simulator for continuous bidirectional associative memories, pretty much as described in bart kosko in several of his articles. i am currently considering fixed autoassociative systems of the form | a1 | | : | | a1 | | I1 | | a2 | | -I : W | | a2 | | I2 | | .. | | : | | .. | | .. | d | aN | | : | S( | aN | ) + | IN | ----| b1 | = | .....:....| | b1 | | J1 | dt | b2 | | : | | b2 | | J2 | | .. | | W' : -I | | .. | | .. | | bN | | : | | bN | | JN | where S() is a sigmoid function to ensure boundedness of synapse signals, Ix is a 'fairly-constant' input to neuron x in one layer, and Jx is an input to neuron y in the other layer. W is formed as W = [ x1 x2 ... xn ]' * [ y1 y2 ... yn ] = X' * Y where X=Y since this is an autoassociative system. I have been simulating the behavior of this system using Matlab, and have not had much luck observing CAM properties in the system. This may be due to a lack of proper understanding of the system operation. For instance, what is the procedure for evoking a memory? Are the inputs to one or both neuron fields clamped? For how long? etc... etc... etc... While I could try a few combinations to figure this out, the simulation time gets pretty high. Does anyone have experience with continuous BAMs or simulating this type of system? Any comments or suggestions would be greatly appreciated. Reply by email to dwatola@eecs.wsu.edu