Both early encoded data and recently trained data contribute similarly to the performance of old/new judgments. When we design and build a computational model for recognition memory suitable for lifelong learning, the properties of familiarity described above need to be considered. In

the following, we survey previous studies on recognition memory at various computational Letrozole price levels. 2.2. Computational Models for Recognition Judgment Recognition memory has been considered a special function of the human brain, rather than a structured type of memory. As a compositional model for the human brain, research into the cognitive architecture has tried to arrange special units for recognition memory. In ACT-R, list memory has an integrated structure that includes recognition memory [22, 23]. In this architecture, recognition memory is depicted

as a simple function occurring in short-term memory and not in long-term memory. The model is unconcerned regarding the SDT or the difference between familiarity and recollection. Based on the ACT-R, a heuristic recognition test was executed for a simple binary judgment [24]. Soar, which is known to be a progressive architecture, judges familiarity according to the success of retrieval in episodic memory [25]. If the retrieval is successful, the input data are regarded as familiar. In this architecture, old/new judgment is not involved. This process considers the recollection process between two items as recognition memory. In the above cognitive architectures, the recognition memory operates as an intelligent function working concurrently with implicit memory and association. However, an independent module for recognition memory is not involved. Recent research has tried to combine the recognition function on Soar. In particular, Li et al. proposed a mathematical approach to reduce the computational cost for searching through long-term memory [26]. This study contributed to the interactional functionality between the

recognition memory and the existing cognitive architectures. Mathematical and computational models for recognition memory have also been studied using the global matching algorithm. SAM [27], MINERVA2 [28], Matrix [29], and TODAM [30] are global matching models that judge familiarity by considering the relationship between a test item Cilengitide and memory [31]. In these models, judgment decision is made quickly, and the SDT is applied to evaluate the performance. The REM model judges the old and new using a Bayesian computation [32]. It computes a scalar value indicating the global matching between the test data and stored memory traces. Cox and Shiffrin advanced the issue of recognition memory by considering the dynamics [33]. According to the data type treated in recognition memory, the criteria for decision of judgment vary in acquiring a constant performance. 2.3.