Implementation of the Hopfield decomposition and DFT into analog hardware
Fisher, Okamura, Smithson, Specht · 1989
Summary form only given, as follows. A 256-neuron, fully interconnected neural net breadboard has been constructed of off-the-shelf components. To date, it has been used to test a 128-node discrete Fourier transform (DFT) feedforward-type net and a 16-neuron Gaussian decomposition, Hopfield-type network. The DFT feedforward net has a settling time of 100 mu s, dominated by the characteristics of the operational amplifier used. The decomposition network had a settling time of 150 mu s, which is 1.5 times the feedforward loop time. A formalism has been identified for setting the optimal local (gain-determining) feedback resistance. Using this technique, 'false minima' solutions are avoided completely. Also, unlike the Tank and Hopfield result, the 'gain' is not run in the high gain limit. This allows the magnitude of the input function to be deduced.>