Serial recall task

國

立 政 治 大 學

‧

N a tio na

l C h engchi U ni ve rs it y 1 . Introduction

Memory is one of the most well-discussed field in cognitive psychology. Researchers in the past one hundred years have devoted to understanding the nature of it, includ-ing how we represent what we experience and how we encode and retrieve them.

Among the various memory phenomena, recall and recognition are the two most studied. Also, there have been proposed a great deal of models respectively for each of these two phenomena. In the past researches, these two kinds of memory were often addressed separately. However, these two kinds of memory might not be com-pletely separated, according to the common understanding about the information processing of human cognition and the evidence implying shared mental process and representation between them. The principal goal of this study is to examine this possibility via computer simulation and modeling. The main hypothesis of this study is that the processes of recall and recognition actually have the same encod-ing process and form the same representation, with the retrieval process as the only difference. In order to verify this hypothesis, a contemporary successful model for serial recall, the SOB (Serial–Order in a Box) model, will be extended to accounting for recognition results. Specifically, in accordance with the hypothesis, the encoding process and the representation generating mechanism in SOB model for serial recall will be remained in the series of modeling. Only will the retrieval process of the SOB model be modified to suit the recognition task. The hypothesis will be supported if the behavioral results of recognition can be accommodated by such a modified model.

1.1 Recall and recognition

1.1.1 Serial recall task

Since Ebbinghaus (1885) studied how well the learned items could be memorized along time, human recall process has been widely investigated. There are various

‧ 國

立 政 治 大 學

‧

N a tio na

l C h engchi U ni ve rs it y

kinds of recall task developed for different purposes, such as free recall, cued recall, forward recall, backward recall and so forth. Among these recall tasks, the serial recall task is one of the most typical case in examining recall performance in STM (Short-Term Memory). Normally, in doing the serial recall task, participants will be given a list of items (e.g., digits, characters, or words) to learn and asked to recall them in their presenting orders after learning. An example of the experimental procedure of serial recall task can be seen in Figure 1.1.

? 9

4 7

+

Response:

７４９

Figure 1.1: The procedure of serial recall task with list length 3.

The performance in the serial recall task is usually measured by the probability of correct (P C) on the items. A response is correct in terms of that the correct item is recalled in the correct position in the memory list. The P C of item typically varies along the item position in the memory list. The P C is higher at the beginning of serial position (the primacy effect), lowest at mid of memory list, and rising at the end of the list (the recency effect), as a U-shape curve in Figure 1.2.

Plenty theories of serial recall are conducted to explain the wealthful phenom-ena found by previous studies. Many theories go into computation model level and proposed different assumptions in detail about representations of memory, encod-ing/retrieval processes, and sources of errors, and their assumptions are support-ed/rejected by carefully designed experiments. The following section describes the major arguments between those models.

The representations of memory trace mainly concerned in models are what

‧ 國

立 政 治 大 學

‧

N a tio na

l C h engchi U ni ve rs it y

1 2 3 4 5 6 7

0.0 0.2 0.4 0.6 0.8 1.0

Serial Position

Proportion Correct

Figure 1.2: The serial position effect in serial recall task. This figure is adapted from the left pannel of Figure 2 in Lewandowsky and Farrell (2008)

participants learded for certain task. In serial recall task, participants have to remember both the items in the trial and the serial order of those items. Different models purpose different nature of how serial position is memorized. Generally, three types of representation are used for memorizing serial position: chaining, position marker, and ordinal.

Chaining models assume that the serial position is coded by the association between neighbor items (Lewandowsky & Murdock, 1989; Elman, 1990). In the chainging model purposed by Lewandowsky and Murdock (1989), current presenting item and the association between the current and the previous item are encoded to the same memory storage. The first and the last item are associated with the start and end signal for each trial, respectively. While recalling, the start signal is served as the retrieval cue to retrieve the first reponse. The retrieved item is then used for retrieving following item as retrieval cue. This cycle continues till the ending signal is retrieved.

Models with position markers propose that serial position is memorized through the association between items and markers with positional information. The

posi-‧

tion marker could be the temporal-based context or event-based context. In OS-CAR (Brown, Hulme, & Preece, 2000), context markers which items associate with are generated by a set of oscillators which gradually change with time. Similiar temporal-based context is adapted in the phonological loop model proposed by Burgess and Hitch (1999) and SIMPLE (Brown, Neath, & Chater, 2002, 2007).

Event-based context, unlike temporal-based context changes with time passing, evolves through event happening. C-SOB (Farrell, 2006; Lewandowsky & Farrell, 2008) and SEM (Henson, 1998) adopte event-based markers. In C-SOB model, the context which associates with item gradually changes with item presenting. There are two contexts mark the start and the end of trial in SEM, and those contexts have similiar role as the context in C-SOB. Regardless the base of context, the retrieval process is quite similar as that the context is treated as retrieval cue for response.

However, the difference in inter-item interval does not influence the prediction of event-based context models but temporal-based context. This difference is exam-ined by the studies around teimpral isolation effect (Morin, Brown, & Lewandowsky, 2010). Temporal-based models predict that the temporal isolated items should per-form better than crowded items. In contract, event-based models predict no tempo-ral isolation effect. Researches do find tempotempo-ral isolation effect in serial recall and suggest that memory trace consists of temporal information (Morin et al., 2010).

Ordinal models suggest that positional informaion is not memorized through item-item associaions or item-context associations but primacy gradient and re-sponse supression (Page & Norris, 1998; Farrell & Lewandowsky, 2002). Primacy gradient is that the encoding strength gradually decreases along with serial position.

While retrieval, the item with hightest activation (eg., first item) is most likely to be responsed. Responsed item is then supressed from memory trace, thus the second item becomes the hightest activated and is most likely to be recalled next. There is no specific representation of serial order in ordinal models, instead the encoding and retrieval mechanism along are enough for reporting items serially.

Primacy gradient and response supression are essential functions in ordinal models and also adopted in many context models. OSCAR , C-SOB, and SEM both have primacy gradient while encoding. However, it is important to note that the

‧ 國

立 政 治 大 學

‧

N a tio na

l C h engchi U ni ve rs it y

course of primacy gradient is only explaimed by SOB (Farrell & Lewandowsky, 2002) and C-SOB (Farrell, 2006; Lewandowsky & Farrell, 2008). In both models, primacy gradient is the consquence of energy-gated encoding. Energy-gated encoding served as that the novelty of incomming information (energy) will modulate the encoding strength of incomming item. The novel items are encoded with greater strength, and the familiar items are encoded lighter. The more items storaged in memory trace, the more overlapping between incomming item and previous storaged memory trace and results in lesser novelty for later items.

Interference from the other items in the memory trace is one of the major source of error in the models previous mentioned. In the chaining model, the re-trieving process will be interference by the non-target chains and items since that every information are stored in the same place. In position marker models, similiar mechanism is applied as that the context-item association crosstalk to each other and cause interference while the context cue is used for retrieval. In ordinal model, like Primacy model, the activation of item is noisy, and the item with highest acti-vation is responsed. The target item, theoretically should be the highest activated item, has to compete with remaining items because of the activation is noisy (Page

& Norris, 1998). The interference of non-target items happens in the deblurring process in SOB model with similiar fasion (Farrell & Lewandowsky, 2002). Besides interference, decay is also a common mechanism that couses error in serial recall.

The activation of item decays over time, and the influence from noise increases along with more decay (Page & Norris, 1998; Burgess & Hitch, 1999). The debate be-tween decay and interference is one of the heavly discussed topic in memory, and the evidence from researches did not converge. Some researches support that for-getting in retrieval is cuased by interference not decay (Oberauer & Lewandowsky, 2008; Lewandowsky & Oberauer, 2009; Brown & Lewandowsky, 2010), and some researches against this (Portrat, Barrouillet, & Camos, 2008).

‧

In recognition, Sternberg’s memory scanning paradigm is the most classical paradigm.

In the original task, participants are successively presented a list of single digits to learn, followed by another digit (i.e., the probe). The participants are asked to judge if the probe was presented in the memory list as quickly as possible (Sternberg, 1966, 1969). The paradigm is shown in Figure 1.3. If the probe is actually presented in the memory list, then it is called the positive probe, otherwise it is called the negative probe.

Figure 1.3: The procedure of recognition task with list length 3.

The basic findings of Sternberg’s task can be concluded by two general laws (Sternberg, 1966, 1969). First, the response time (RT) is linearly increasing with the set size. The zero-intercept of this linear regression line is 397.2ms, and the slope is 37.9 ms for each additional item in memory list. Second, the slope of RT for “yes” response is equal to “no” response. The result is shown in Figure 1.4. The similar results are replicated with characters, artificial symbols, words, and colors (Sternberg, 1975).

Sternberg explains this finding with his exhaustive serial search model (Sternberg, 1966, 1969). In his model, the items in the memory list are marked with special markers serially during initial learning. When determining if the probe has been presented or not, a high-speed exhaustive scanning process compares the item in the list with the probe serially. If a match is found, then response is “yes”, and response is “no” otherwise. However, the scanning process does not stop even when