Dual-memory retrieval efficiency after practice: effects of strategy manipulations

Abstract

The study investigated practice effects, instruction manipulations, and the associated cognitive architecture of dual-memory retrieval from a single cue. In two experiments, we tested predictions about the presence of learned parallelism in dual-memory retrieval within the framework of the set-cue bottleneck model. Both experiments included three experimental laboratory sessions and involved computerized assessments of dual-memory retrieval performance with strategy instruction manipulations. In Experiment 1, subjects were assigned to three distinct dual-task practice instruction groups: (1) a neutral instruction group without a specific direction on how to solve the task (i.e., neutral instruction), (2) an instruction to synchronize the responses (i.e., synchronize instruction), and (3) an instruction to use a sequential response style (i.e., immediate instruction). Results indicate that strategy instructions are able to effectively influence dual retrieval during practice. Mainly, the instruction to synchronize responses led to the presence of learned retrieval parallelism. Experiment 2 provided an assessment of the cognitive processing architecture of dual-memory retrieval. The results provide support for the presence of a structural bottleneck that cannot be eliminated by extensive practice and instruction manipulations. Further results are discussed with respect to the set-cue bottleneck model.

Appendix

The set-cue bottleneck model of dual-memory retrieval

The set-cue bottleneck model assumes three sequential, independent, and additive processing stages: perceptual, central, and motor processing. Figure 11 shows how these three stages map onto the set-cue architecture to be described in detail next. The perceptual stage includes all activation flow from the moment of cue presentation up to threshold-level activation at the input level. The retrieval stage spans from the moment of activation gating (i.e., the moment of threshold activation) from input level to threshold activation at the response level. The motor stage includes the gated activation flow from the response level to completion of the motor response. These three stages can be mapped directly to the perceptual, retrieval, and motor stages of the set-cue bottleneck model.

Fig. 11

Basic architecture of the set-cue bottleneck model including perceptual, retrieval, and motor stages. Nodes at the input level represent the presented cue and the currently active task set(s); each node at the set-cue level represents a particular combination of presented cue and task set; each task set node in turn can have an association of one or more nodes at the response level. K = task set of keypress task; V = task set of vocal task

The set-cue bottleneck model assumes that associations are formed only between activated nodes at adjacent levels of the hierarchy. Nodes at the input level represent the presented cue and the currently active task set(s). A task set will be defined for current purposes as a goal to execute one of a set of responses in a given modality. For the keypress task, the task set node (K) represents the goal to execute a correct key press (left or right) when the cue is presented. The vocal-digit task set node (V) represents the goal to execute the correct vocal-digit response when the cue is presented. Activation of task set nodes is, thus, assumed to be under strategic control. Note that for a given task (keypress or vocal) there is only one task set node.

Each node at the second level of the set-cue bottleneck model (the set-cue level) represents a particular combination of presented cue and task set. Each task set node in turn can have an association with one or more nodes at the response level. Nino and Rickard (2003) proposed that, whereas activation streams can flow in parallel from the earliest level of cue perception to the set-cue level, and from the set-cue level to the response and motor levels, there is a winner-take-all competition at the set-cue level. After a winning set-cue node is selected, activation begins to flow from (only) the winning set-cue node to the response level. In the model as developed here, the entire retrieval stage of processing—from the moment at which activation flow is gated forward from the input level to the moment of threshold activation at the response level—must be completed before activation flow for a second retrieval event can be initiated from the input level.

To specify mechanisms of learned parallelism in grouper subjects, snapshot activation states of the set-cue bottleneck model during dual retrieval are shown for two cases in Fig. 12. The first case depicts dual retrieval for groupers on the first dual-retrieval trial. Panel A depicts the node activation state just after cue presentation and just prior to activation gating to the set-cue level. Because both responses are required on dual-retrieval trials, both task sets are assumed to be activated at the outset of each trial. Activation then flows in parallel to all associated nodes at the set-cue level. The example case in which the set-cue node corresponding to the keypress task (labeled as node 1) wins the competition (i.e., reaches an activation threshold first, resulting in suppression of activation in all other nodes at that level), is shown in Panel B. Factors that could determine the winning set-cue node include noise, differential strength of the associations formed during single-retrieval phase of learning, or strategic scheduling that gives preference to one category of retrieval (i.e., keypress or vocal-digit) for first execution. Strategic scheduling could be implemented within this framework as enhanced initial activation of one of the task set nodes.

Fig. 12

Snapshot activations of dual retrieval in the set-cue bottleneck model for nongrouper subjects throughout dual-retrieval practice (first row) and grouper subjects (i.e., groupers) at the beginning of dual-retrieval practice (second row) and with dual-retrieval practice (third row)

After the winning set-cue node reaches an activation threshold, activation then flows to the response level (Panel C). Because groupers are synchronizing their response execution, however, they do not immediately execute the first retrieved response, but rather keep that response active by a working memory mechanism as they execute the second retrieval. It is also reasonable to assume that the task set corresponding to the first retrieved response also remains activated until that response is executed. For set-cue node 2 to win the second competition, some mechanism that biases the competition in favor of node 2 must be assumed. Given such a mechanism, set-cue node 2 would win the second competition, and activation would flow from that set-cue node to the vocal-digit response (Panel D). At that point, both task sets and both responses are activated and there is synchronized keypress and vocal-digit response execution.

Important for the present context, we assume that a necessary and sufficient condition for associations between two nodes to be formed or strengthened is their joint activation. It follows that, during the activation state depicted in Panel D, there will be strengthening of the existing associations as well as formation of new associations between a) the keypress task set and the active set-cue node 2, and b) set-cue node 2 and the active keypress response node, as depicted in Panel E under the heading “Grouper subjects (with dual-retrieval practice)”. In the set-cue bottleneck model, then, there is no task-level learning associated with response chunking or learned retrieval parallelism. Rather, response chunking occurs independently for each cue (i.e., cue-specific).

After one or more grouped response trials for a cue, the associations represented by the dashed arrows in Panel E become strong enough that they can support retrieval of both the vocal and keypress response via set-cue node 2. On subsequent dual-retrieval trials, set-cue node 2 is likely to win the initial activation competition at the set-cue level even though it was not the initial winner on previous dual-retrieval trials for that cue (in this example). This outcome would be expected because, at that point during training, both task sets would send activation to set-cue node 2, whereas only the keypress task set would send activation to set-cue node 1. Furthermore, because set-cue node 2 is linked with both response nodes, it activates both a keypress and a vocal-digit response. Hence, subjects can learn to retrieve both responses while making only one pass through the set-cue bottleneck stage, i.e., subjects can learn to chunk responses. The sequence of activation states for this case is shown in Fig. 12f, g.