We classified three types of edge configurations. The first edge has a fixed order size within the range of 2 to d − 1, where d is the number of dimension. The second configuration sets random order k, which is determined to be between r1 and r2, where

r1 and r2 are in the range BRL-15572 of 2 to d − 1. In the above two edge configurations, the node combination is sequential in order of the number of attributes. Under the assumption of a random edge order, the third type of edge randomly selects order combinations. We applied a total of 13 edge categories, which include fixed-order edges, random-order edges, and random-edge combinations. Next, we investigated the incremental trend of recognition memory according to the amount of encoded data. Eight categories of edge configurations were compared. Figure 8 shows comparison of the encoded memories in terms of the number of edges and links. Additionally, the ratio of links over edges, which is defined as connectivity, indicates the degree of memory density. For a recognition judgment, we expected the connectivity to play a critical role. When fixed-order edges are applied, the scale of the encoded memory increases, and the connectivity decreases. Both random-order edges and random-edge combination also showed this same tendency. However, the average connectivity achieved the maximum value

in the case of random-order edges. Figure 8 Scale of encoded memory and the memory connectivity according to the edge configuration. 4.2.1. Familiarity Judgment In the experiments, we were mainly concerned with finding the evidence indicating that the proposed model has a similar recognition judgment performance as a human being. The Reality Mining dataset is appropriate to represent human behavior because it contains repeated similar events and changes the

data distribution by time. We know that an optimal ROC curve has a high hit rate and low false alarm rate. However, human behavior has uncertainty with false alarms and false negatives. According to the various edge configurations, in this experiment, we drew Cilengitide ROC curves and investigated their properties to search for a human-like configuration. To draw a ROC curve, a similarity measure was applied. Even when a familiarity judgment is executed using the activation-based mechanism, we can acquire a similarity of the activated link using the weight-based mechanism. When the recognition memory uses only the activation-based mechanism, an input instance with a high similarity value can be judged as new. On the other hand, an input with a low similarity value is judged as old if all of the activated edges are connected with each other. For a familiarity judgment, we ignore this situation because we need to obtain quantified data for the ROC curve. ROC curves for the three edge configurations with an order range between 2 and 6 are shown in Figure 9.