A Bottleneck in Memory Retrieval from a Single Cue

We explored whether two memory retrievals from a single cue can be completed in parallel. Subjects were trained to make a vocal response and a key press response for each of a set of ten visually presented words. Subsequently, these two tasks were interleaved with a third, dual task condition in which subjects retrieved both the vocal and the key press response on each trial. Findings include: 1) the latency of the first retrieval on dual retrieval trials exceeded that of either of the two tasks performed by themselves, 2) the time to finish the dual retrieval task was roughly the sum of the time to finish the two tasks on separate trials, 3) the correlation between time to finish the first task (RT1) and the inter-response interval (IRI; latency between the first and the second response) was not significant, and 4) some subjects always choose to execute the slower task first on dual task trials. Taken together, these and other findings favor a sequential retrieval model over a broad class of parallel models. One goal of cognition research is discovery of principles for predicting when two or more processes can or cannot be executed in parallel. This paper examines the case of memory retrieval. More specifically, the paper explores whether two independent memory retrievals from a single cue can be completed, from stimulus perception to response execution, in parallel. This question is fundamental to the study of memory and attention. Yet, remarkably, it appears that no study has ever directly addressed it. Our goals are to document the basic empirical phenomena, and to test candidate theories. Several studies in the literature may provide insight into this issue. Rohrer, Pashler & Etchegaray (1998) investigated category exemplar retrieval in the case in which all items to be retrieved are members of a single category, compared to the case in which subjects must alternate retrievals between two categories. In Experiment 1 of their study, subjects studied four exemplars from each of two categories (e.g., for the category animal, they studied the words goat, horse, pig, and sheep). On some trials, subjects were instructed to recall the examplars from only one of the categories. On other trials, they were instructed to alternately recall exemplars from each of the two categories. The IRIs were much slower on trials on which recall alternated between categories than on Send correspondence to: Timothy Rickard Department of Psychology, 0109 9500 Gilman Drive University of California, San Diego San Diego, Ca 92093 – 0109 [email protected] trials on which recall occurred only within a single category. Based on these and related results, the authors argued that items from different categories cannot be retrieved in parallel. Their conclusion regarding within-category recall, however, was quite different. IRIs progressively increased from the first to the last exemplar recalled within a given trial, and this effect was greater in the alternating category than in the single category condition. The authors showed mathematically that this effect is can be explained if one assumes that withincategory retrieval is parallel, though they did not conclusively eliminate other possible accounts. The between-category condition in Rohrer et al. (1998) can be seen as analogous to a dual task design, since two separate cues (one corresponding to each category label) were presented at the outset of each trial. There is converging evidence from dual task studies employing the psychological refractory period paradigm (Welford, 1952) that parallel retrieval is not possible when there are two separate cues on each trial (Carrier & Pashler, 1995; Anderson and Lebiere, 1998). In contrast, the within-category trials of Rohrer et al. can be seen as somewhat analogous to the design of our current experiment, in which only a single cue (e.g., a single category label) is presented, and two or more retrievals are performed using that cue. Their results suggest that parallel retrieval may be observed in the current experiments. Note, however, that being of the same category, retrievals in the within category condition were not independent in the sense that they are in the current study The results of a study by Ross and Anderson (1981) provide additional evidence Retrieval Bottleneck 2 suggesting that two memory retrievals from a single cue can be completed in parallel. They had subjects study lists of paired associates. Across items in the list, each cue word appeared several times, in each case paired with one of two response words (e.g., baker-garage, or baker-canyon). For all cue words, one of the response pairings occurred frequently in the list (the strongly associated response), and the other occurred infrequently (the weakly associated response). In a subsequent test phase, subjects saw the cue words presented one at a time, with instructions to recall the first response word that came to mind. Not surprisingly, the strong associates were retrieved more often and more quickly than the weak associates. Of more interest, the upper tails of the response time (RT) distributions for strongly and weakly associated response words appeared to converge, a phenomenon identified previously by Townsend (1974) as consistent with parallel retrieval. In their task, however, subjects were only required to complete one retrieval per trial. Their results thus do not bear directly on the question of whether two retrievals can run to completion and be executed in parallel. Their results suggest only that, in their task, both responses were activated in parallel prior to execution of the first response. Although both the Rohrer et al. (1998) and the Ross and Anderson (1981) papers suggest that retrieval of two responses from a single cue might be completed in parallel, recent studies by Rickard (1997; 1999) suggest the opposite. Rickard showed that the strategy shift from use of multi-step algorithms to direct memory retrieval in tasks like mental arithmetic can be naturally accounted for assuming that only one of these strategies can be executed to completion on a given performance trial (cf. Logan, 1988; Nosofsky & Palmeri, 1997; Palmeri, 1997; 1999). The model that he proposed assumes that only one response can be retrieved per attempt and also that most multi-step algorithms used in skill learning tasks reflect one or more longterm memory retrieval steps. It follows from that model that two strategies cannot be executed in parallel. It also follows that two responses cannot be retrieved in parallel given a single cue, or given two different cues for that matter. Thus, there is a theoretical tension in the literature regarding the case of retrieval from a single cue. The findings of Rohrer et al. (1998) and Ross and Anderson (1981) indicate that retrieval of two responses might be completed in parallel. The Rickard (1997; 1999) studies suggest the opposite. This paper is intended to help resolve this tension. In experiment 1, subjects were first trained to make a vocal response (saying a single digit) and a key press response for each of a set of ten visually presented words. The ten word stimuli were mapped to ten vocal (digit) responses, and to two (left or right) key press responses. Subjects were then required to perform a dual task in which they made both the vocal and the key press response on each trial (single-task blocks were interleaved with the dual-task blocks in this phase of the experiment).