1. Introduction
1.3 Computational modeling
國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
1.3 Computational modeling
In order to examine the above idea, computational modeling would be the best way to go. The most typical instance about the implication of computational modeling in psychology comes from the seminal study of Nosofsky (1986). Before his study, the identification task and categorization task were usually considered as two sep-arated tasks. That is, the mental processes as well as the mental representation of identification and categorization were thought to be different. In identification, participants are given a list of learned items and asked to identify them. In catego-rization, participants are given a list of items and asked to classify them to different categories. Intuitively, the regime of these two tasks are quite similar that implies some cognitive components might be shared between them. Nosofsky (1986) veri-fied this idea using his categorization model, the GCM (General Context Model).
According to this model, the stimuli are encoded as exemplars in our mind which in turn are used for categorization. With differential attention allocation on stimulus dimensions, the GCM was evident to be able to account for both the performance in the identification task as well as the categorization task.
The GCM model assumes that an item would be classified to be the category whose exemplars are much similar to the item. The similarity is transferred from the distance between the items in the psychological space. In addition, the distance on each dimension is weighted by attention weight. For instance, if a dimension is more valid for correct categorization, it will be more attended to, whereas the other dimensions will be less attended to. Therefore, if we regard the identification task as a categorization task with multiple target categories instead of just two or one category in the categorization task, then identification can have a great deal of parts overlapped with categorization. That is, to identify an item is equivalent to examining how similar this item is to the learned exemplar, which is actually itself for successful identification.
In the study of Nosofsky (1986), participants were asked to do both the identifi-cation and categorization tasks. In the categorization task, all stimuli were composed of two features, one of which was valid for categorization. The GCM was evident
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
being able to account for the performance in both tasks, with different attention weights on dimensions. That is, for accounting for categorization performance, at-tention weight on the valid dimension was extremely high, however, for accounting for identification performance, attention weights on both dimensions were almost the same. Thus, this modeling result shows that with the same set of exemplars (i.e., representation), plus different attention weights on stimulus dimensions, the two tasks previously thought different are actually quite similar to each other. This is the instance for the contribution of modeling in psychology.
Following Nosofsky (1986)’s methodology, if serial recall and recognition share the same encoding process and representation, there must exist a mathematical model which can account for performance of both serial recall and recognition task well. For this aim, the present models for serial recall and recognition will be discussed to see whether any one of them can actually be the candidate model to examine the hypothesis in this study.
Previous researches proposed several models which could account for recall and recognition. For instance, Search of Associative Memory (SAM) model can account for recall and recognition in long–term memory. Though many phenomena can be accounted for, but not the serial position effect (Raaijmakers & Shiffrin, 1981;
Gillund & Shiffrin, 1984). Retrieving Effectively from Memory (REM) model and its extension ARC–REM could also account for recognition and cued recall in long–
term memory. However, the serial position effect could not be simulated (Diller, Nobel, & Shiffrin, 2001; Shiffrin & Steyvers, 1997) still. Therefore, those models concerning performance related to long–term memory should be excluded from the candidate list.
In short–term memory, there is one connectionism model (see Burgess & Hitch, 1999) could account for short–term memory serial recall and recognition task. This model is developed based on the association between time and item. However, this model shows smaller primacy effect than the observed data in the serial recall task.
In recognition task, the model shows no primacy effect, but recency effect on P C.
RT of recognition cannot be simulated either. Therefore, this model is evident
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
unable to accommodate both serial recall and recognition performance. SIMPLE (Brown et al., 2002, 2007) is also used to model the result of recognition tasks which use unfamiliar faces as stimulus and fits well to the data in experiments (Hay, Smyth, Hitch, & Horton, 2007). However, SIMPLE model does not adress on the encoding and retrieval processes but the representation of memory. The model that I am looking for is the one which describes the process of encoding and retrieval in enough detail. SIMPLE model does not fit the criterior here.
The remaining models in the candidate list can not account both recall and recognition. Therefore, an extension for previous model is needed for previous model to account both tasks. The following concern is which model should be extended.
Many models had been briefly introduced in the previous section. In this thesis, the chosen models are SOB and its successor C-SOB. There are two major reasons for selecting those models: the difficulty of modification and examination of the neccessity of context.
The difficulty of modification to extend SOB and C-SOB into recognition model is low because of the energy in both models. The energy is one of the core mechanism in both SOB model and is defined as the familiarity of incoming item (Farrell &
Lewandowsky, 2002; Lewandowsky & Farrell, 2008). The recognition process could be defined as the judgement of familiarity. SOB and C-SOB model provide a build-in mechanism to determbuild-ine the familiarity of probe. This reduces the difficulty to extend persist model dramatically.
Second reason for chosing SOB and C-SOB is that the major difference between SOB and C-SOB is the context layer. As mentioned above, SOB model is an ordinal model which encodes order information through signal strength, and serial position is represented through context marker in C-SOB. Hitch, Chiara Fastame, and Flude (2005) used Hebb procedure to examine the underlying mechanism and suggested that the position information plays an important role in serial recall. In other hand, the context information might not be needed in recognition, especially in Sternberg’s task (Corbin & Marquer, 2008, 2009). The necessety of context marker in recognition is examined by comparing the simulation performance between SOB
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
and C-SOB.