| 摘要 |
<p>A method and apparatus for implementing a neural network having a sleep mode during which capacitively stored synaptic connectivity weights are refresehed. Each neuron (16-1...16-3; 18-4...18-7) outputs an activity level represented by a pulse frequency. The apparatus includes feed-forward (30-14...30-17) and feed-back (32-41...32-71) synaptic connection circuits whose respective couplings are weighted according to values stored on capacitors in each connection circuit. A learning algorithm strengthens connections in both directions between particular ones of first level (16-1...16-3) and second level neurons (18-4...18-7) to the extent of concurrence of high activity levels in both, and weakens same to the extent of concurrence of high activity level in the second level neuron (18-4...18-7) and a low activity level in the first level neuron (16-1...16-3). A sleep mode disconnects all environmental inputs (22-1...22-3) and provides non-specific low activity level signals to each first level neuron (16-1...16-3), causing the network to randomly traverse its state space and accomplishing refresh as the learning algorithm remains active. A deep sleep mode prevents runaway strengthening of favored states and encourages Weber law compliance.</p> |
