Autoassociative neural memory capacity and dynamics

A new synthesis approach for feedback neural networks based on the perceptron training algorithm

“…A more sophisticated definition of associative memory capacity has also been used in the literature [8], [9]. The capacity used in [8] and [9] is defined as a measure of the ability of an associative memory to store a set of unbiased random binary patterns at a given error correction and recall accuracy level (e.g., 99%). Using this definition of capacity, the present method and the Ho-Kashyap recording [8], [9] will be compared (cf.…”

“…The study of such systems has been of great interest to many researchers in recent years (see, e.g., [1], [4]- [9], [11]- [15], [17], [21]- [23], [26]- [29], [31], and [33]). These works are concerned with the qualitative analysis of neural networks [1], [4], [5], [7]- [9], [11]- [15], [17], [22], [23], [31], and design methodologies for such networks [6], [8], [9], [11], [14], [15], [17], [21]- [23], [26]- [29], [33]. The study of singlelayer sparsely interconnected feedback neural networks has also been of recent interest [2], [3], [17]- [20], [24], [32].…”

“…In these synthesis methods, a set of linear equations is formulated and solved for the design of neural networks. In the design method advanced in [8] and [9], a set of linear inequalities is formulated and solved for the optimal meansquare-error (MSE) solution using the Ho-Kashyap method [10]. In the design method developed in [27], a set of linear inequalities is formulated and solved using optimization techniques.…”

Associative Memory Design Using Support Vector Machines

“…The value corresponds to ; taking into account that there are 45 vectors at , the recall accuracy is 86% with the SVM and 58% with the perceptron. For sake of comparison we consider the results in [17], based on the Ho-Kashyap method [18]: They are % for and % for and . Example 2: Here, we consider a specific design example and compare four different design strategies: SVM standard, SVM with fixed threshold according to (23), perceptron, and the designer neural network of Chan and Zak.…”

“…Note that a subset of the Lagrange multipliers will be zero (corresponding to non-support vectors), hence not all the products will be present in (16). Often the design of model (12) is formulated as follows [6], [7]: Find such that (17) where is called margin of stability. Since the weight vector can be rescaled to obtain any value for , a limitation or normalization of the weights is required in this formulation (e.g., ).…”