Layered dynamic auto-associative memory with auto-encoder and feedback

Summary form only given, as follows. An aim of this study is to propose a layered dynamic autoassociative memory which treats binary patterns as input/output and consists of a (multi-)layered network trained by the backpropagation and a simple feedback from output to input. This simple-architecture, autoencoder-type feedforward network with feedback dynamics works as an extended version of autoassociative memory and overcomes the poor ability of the autoencoder for recalling a learned pattern. It also solves two critical problems of conventional correlation-type associative models: crosstalk noise elimination and multiple-match resolution. It is shown that decision hyperplanes, which are formed in a multilayer feedforward network by means of the backpropagation or LMS (least-mean square) learning algorithm, characterized the performance of the proposed model.<>