Hybrid location-content addressable memory
Bixuan Wang, Seak Bai Koh, Seung Kwon Ahn · 2002
A neural network model, called the hybrid location-content addressable memory (HyLCAM), for representing binary-to-binary mappings is presented. A cascaded connection of a single-layer perceptron and a location addressable memory characterizes the structure of the HyLCAM that implements the concept of indirect association. A three-tuple (X,Z,Y) defines the state of the HyLCAM. The intermediate state Z plays an important role in learning and storing the given association (X, Y). The key of HyLCAM encoding is that a designer is to construct the immediate states so that they are linearly separable with respect to inputs. This manual operation, referred to as code generation, eliminates the limitations of back propagation, such as a slow learning speed and the convergence to error local minima. Two simple code generation methods are developed, and their performances are compared.>