The objective of this thesis is to provide various types of models and architectures for bidirectional associative memory(BAM) that are fast, efficient, and convergent. By reflecting feasible concepts which are abstracted on the analogy of the human brain, generalization of previous models and architectures is performed. First, weighting matrixes are introduced to input and output side of the original BAM model to obtain input-weighted BAM (IWBAM) and output-weighted BAM(OWBAM) respectively. For both models, learning algorithms are proposed, and theorems for existence and convergence of learning parameters are also derived. Second, the proposed models are further generalized by nonlinearization to overcome the generic storage capacity and error-correction capability limitations of linear models, and nonlinearized IWBAM(N-IWBAM) and nonlinearized OWBAM(N-OWBAM) are obtained. These models are endowed with learning algorithms, characteristics of which are analytically investigated through statistics. For the N-OWBAM, existence and convergence theorems are induced similarly to OWBAM. Using the proposed models which can discriminate highly correlated patterns, a scheme for multiple association is developed to store and recall spatio-temporal sequences robustly within the framework of Wang`s architecture. Computer simulation is performed to validate theoretical results and to compare the proposed models with other BAMs with respect to fundamental properties of BAM such as storage capacity, noise immunity, spurious memory, etc.