Showing papers on "Bidirectional associative memory published in 1985"

Virtual uteral inhibition in parallel activation models of associative memory

Abstract: This paper describes a new theory of how spreading activation may occur in associative memory models formulated as parallel activation networks The theory postulates that competition for activation by nodes/concepts in a network is a fundamental principle of memory retrieval Using only excitatory connections between concepts, a specific implementation of this model is able to demonstrate "virtual lateral inhibition" between competitors and other interesting behaviors that have required use of explicit inhibitory connections in the past

An associative memory model based on adaptive feature-detecting cells

Abstract: This paper extends the function of Amari-Takeuchi's model by adding an inhibitory cell with nonlinear characteristics to their self-organizing model of the feature-detecting cell. The generalized model is shown to separate even patterns in mutually inclusive relation, forming the feature-detecting cells for each of the patterns. It is shown that when a learning parameter in the model is modified, it can perform classifications under various criteria based on the similarity. Then an associative model based on the feature-detecting cell is proposed. The model is composed of three subsystems: the feature-detecting cell formation subsystem to convert in a self-organizing way the input pattern to the specified feature set; the learning subsystem, which associates in a multiple way two kinds of signals by cross-correlation; and the decoding learning subsystem, which restores the original pattern from the coded signal. The behavior of the model was examined under a multiple-associative environment. The model can handle the mutual and self recalls in a unified way, and the memorized patterns associated with the key pattern are recalled successively. The result indicates a role of the feature-detecting cell in the associative memory. It is also suggested that the model can be a basis for accounting for the high-level associative function through association of high-level notions.

Two-stepped association procedure and its application to reduction of memory areas of associative memory

Abstract: This paper proposes a new association scheme (called two-stepped association scheme) which executes the traditional distributed association scheme in two steps, analyzing its association ability. As an application of this association scheme, a method in the distributed associative memory is proposed which can reduce optimally the size of the memory area for memorizing information to the sense of the least-mean-square error. In the two-stepped association scheme, two-step associations are performed, which are from the key to the mediate vector and from the mediate vector to the data. The associative process in the associative memory can be made flexible by a suitable selection of the mediate vector. One advantage of the two-stepped association scheme is that the model acquires a larger degree of freedom in its structure by executing the associative process in two steps. In this paper, this point is utilized and a discussion is made for the condition and the determination of the mediate vector, which leads to the construction of the associative memory with the memory area of specified size and the least mean-square-error between the association output and the data. It is verified by a computer simulation that the associative memory constructed in this way can realize the desired associative ability.