Skip to Main Content
Recurrent-convergence decision circuits include the historic Hopfield neural networks as well as prospective turbo decoders. For these applications, we propose two novel circuit elements: (1) a quick-response synaptic weight with high symmetry in both small-signal and large-amplitude operations and (2) an open-loop, quick-response and symmetrically parallel neuron circuit that has inverse and non-inverse outputs with delay-time equality and high signal-amplitude symmetry. We report the characteristics of these elements in a fabricated associative memory chip. We also analyze how the recalling statistics depend on the variation of the symmetry and delay equality of the elements. It is demonstrated that the proposals are very effective to improve the neural association performance.