Associative amplitude modulation with built-in noise immunity

Single-pattern real-valued spectral associative memories (SAMs) are proposed for coding and recalling sampled analog or multi-valued data patterns over noisy channels. SAMs are frequency-domain formulations of associative memory that combine the extrinsic redundancy of neural networks with the in-phase and quadrature modulation schemes of telecommunications. Data patterns, or "codewords", are encoded into an attractor wave by associative amplitude modulation (AAM), which may be recalled by recurrent associative amplitude demodulation (AAD). Since no attractors are actually stored in the decoder, single patterns may be transmitted, one at a time over a noisy channel with spurious-free recall. Long-range connectivity is made virtually in the frequency domain, allowing both encoder and decoder to scale linearly with pattern dimension. In-phase and quadrature coding schemes are presented with band structures and anti-aliasing constraints for auto- and heteroassociative memory formation and recall. Simulation results are provided that show the accuracy of recall for various Butterworth and Chebyshev filters.