| Abstract | Capacity is an important factor in the concept of memory. Bidirectional Associative Memory (BAM), one model of associative memory which also belongs to a class of neural networks, has aroused much interest because of its potential benefits. A promising model, its information processing mechanism needs investigations. It is desirable that BAM should have as large a capacity as possible, i.e., the ability to store and recall properly a large number of pattern pairs. It storage capacity has been estimated at C< min (n,p). where n and p are the dimensions of pattern pairs A and B respectively. The capacity is somewhat peculiar in that the network can recover only C memories out of the total 2n and 2p states available in the network as cube corners of n- and p- dimensional hypercubes for patterns A and B respectively. Such a capacity can be considered very low. Two issues are addressed in this study. Firstly, it investigates some important properties of BAM and proposes an improved capacity estimate. Those properties are the encoding form of the input pattern pairs as well as their decoding, the orthogonality of the pattern pairs, the similarity of associated patterns, and the density of the pattern pairs. Secondly, it proposes an implementation approach to improve the storage capacity. The approach embraces three proposed methods, i.e. the Bipolar-orthogonal Augmentation, the Set Partition, and the combined method. Along with those proposed methods is the construction of the set of bipolar orthogonal patterns. Four important properties of BAM have been investigated and analyzed. First, bipolar encoding-decoding gives the best performance over the other possible encoding-decoding forms. Second, for a trivial case is that BAM's capacity will be bounded at C= min (n,p) when all the pattern pairs are orthogonal. When they are non-orthogonal, the capacity decreases. Third, for another trivial case where the key pattern is similar to the associated pattern, i.e. A=B and thus n=p shows that BAM can store and properly recall all 2n available states of the pattern pairs if it is encoded in bipolar form. Fourth, patterns encoded in a balanced density give better recall operation than the other encoding. The implementation approach consists of modification of BAM's original architecture, and handing of input pattern pairs. The modification of BAM's original architecture includes modification of related BAM algorithms. Two methods, Set Partition and Bipolar-orthogonal Augmentation are proposed where each of them guarantees the recall of all pattern pairs. Both methods can be combined to produce optimum performance. For the purpose of the approach, an algorithm to construct a set of bipolar-orthogonal partners has been developed. Sufficient conditions for the length of augmentation for which all of the patten pairs can be stored and properly recalled in BAM have been derived. The trade-offs between the improvement of the capacity and the memory complexity are also discussed. |