We selected Bayesian networks (BN) to infer the missing part from the partial data. If the contextual data is built with probabilistic distribution table, the model can expect the related event from partial instance. The Bayesian network was also tested using the Reality Mining data. To infer an event through the BN model, structure learning and parameter learning are required. Each value in the data mTOR inhibition is composed of categorical data, so that we used an algorithm from Auton Lab [47]. The parameter learning was executed by using commercial BN product. Similar to offline hypernetworks, every 1000 instances

are used to update the BN model. Then the next 1000 instances are tested whether the model expects the missing values well. Hence, the expectation starts at 1000th instance. At first, the expectation was higher than other memory models. However, the performance decreases by time and the final performance was 13%. The probabilistic model is hard to keep the less probable events. The probabilistic distribution table extracts the most probable values from the conditional probability. This experiment shows that that online hypernetwork is more adaptable than the probabilistic approaches for pattern

completion and expectation in lifelong experience. 5. Discussion 5.1. Tradeoff in Performance Based on the Connectivity We evaluated the proposed recognition memory model in terms of familiarity. We investigated two functionalities of recognition memory, old/new judgment as explicit memory, and pattern completion as implicit memory. From the various edge configurations, we found a tradeoff in the two functionalities.

For old/new judgment, we searched the optimal conditions for a hypergraph structure that resembles the recognition memory based on human behavior. If the memory model merely acts as a judgment model, the memory model should separate old and new instances perfectly. When we model the memory with a high number of fixed-order edges, we can reach the memory goal. However, old/new judgment is an explicit function of recognition memory and only works for complete input data without AV-951 missing values. Additionally, we focused on another characteristic of recognition memory, that is, the implicit function. When partial data with missing values are assigned as an input value to the encoded memory, the performance is not indicated by the ROC curves, which deals with true and false positives, but by the possibility to generate the original complete data. We found that the explicit and implicit functions have a tradeoff relationship, and thus we need to select the optimal conditions for those two distinguishable processes. The main criterion for the performance was network connectivity in the memory model. The memory model revealed a different connectivity according to the edge configuration. A model with a large number of fixed-order edges has a tradeoff relationship with a model with a small number of fixed-order edges.