Serial Order Memory, Computational Perspectives

Representing Order

One basic issue addressed by these models is how the order of elements in a sequence is represented. Figure 1 shows three general schemes of representation of order in contemporary models of serial order. In the top scheme, chaining, adjacent elements of a sequence are associated in memory. Once an item is recalled, the following item can be accessed by using the recalled item to cue the next item via the pair association. Although Stephan Lewandowsky and Bennet Murdock showed that the chaining model could account for many of the key phenomena described above, later work has challenged this model. A particularly troublesome finding is that alternating rhyming and nonrhyming items in a sequence does not harm—and may even enhance—recall of the dissimilar items. Chaining models predict that confusions of rhyming items should be followed by confusions of nonrhyming items: When a rhyming item is recalled in the wrong position, the following nonrhyming item should move with it and thus be incorrectly ordered.

Figure 1 Schemes for representing order in memory

This led to the development of primacy gradient models, in which order is represented as a gradient of activation or encoding strength across items. These models successfully account for many of the key phenomena and can allow nonrhyming items to be protected from confusions between rhyming items by assuming that phonological confusions [Page 686]occur in a stage downstream from the primacy gradient. One challenge to these models is the effect of grouping on serial recall: Temporally grouping a sequence by placing pauses between subsequences produces primacy and recency within subsequences and also leads to specific patterns of confusions in between groups, the most telling of these being that elements that appear at the same position in different subsequences are more likely to be confused. These grouping effects imply some form of factorial or hierarchical representation in short-term memory, taking these data beyond the limits of the unidimensional, strength-based, primacy-gradient models.

...