BAM: A Bidirectional Associative Memory Neural Network Architecture

BAM (Bidirectional Associative Memory) is a two-layer hetero-associative neural network architecture forming a feedback system built from content-addressable memory components. The network supports multiple variants including discrete, continuous, and adaptive BAM implementations. Originally developed by Kosko in 1987 and 1988, BAM operates bidirectionally: when an input signal is applied to one layer, it generates corresponding output at the opposite layer. This output then feeds back through the network, undergoing multiple iterative cycles until the system reaches equilibrium. From an implementation perspective, BAM typically employs a weight matrix learned through Hebbian or gradient-based learning rules. The discrete version uses binary threshold activation functions, while continuous variants apply sigmoidal functions for gradual transitions. Key algorithmic steps involve matrix multiplication between input vectors and weight matrices, followed by activation function application and iterative feedback processing. BAM finds extensive applications in pattern recognition and information processing systems, particularly for bidirectional pattern association tasks where incomplete or noisy inputs can recall complete stored patterns through the network's convergence properties.