The replacement effect: Repeating some items while replacing others

Abstract

We performed three experiments on recognition learning that tested for the existence of a replacement effect (i.e., the benefit accruing to nonrecognized items, or targets, when recognized items are replaced in the next study trial). A reverse Rock substitution procedure was used, and the replacement effect occurred in all three experiments. The results were interpreted in terms of a distributed memory model, the matched-filter model of Anderson (1973), but several modifications were necessary. The original version cannot learn, and a closed-loop modification did not show the repetition effect that was clearly evident in the data. The most satisfactory version was one based on probabilistic encoding of features in the item vectors, and it seemed capable of explaining most aspects of the data.Recent developments in distributed memory models provide a new way of looking at human memory. They specify how information can be stored and retrieved, they explain how recognition and recall can occur, and, coupled with a decision system, they can account for both accuracy and latency data. They are content addressable, and retrieval occurs by direct access rather than by search. They are consistent with known physiological mechanisms; in fact, many distributed memory models (see Hinton & Anderson, 1981) are based on or are outgrowths from neural models.Originally, it was not clear how to test distribut¢ memory models, so their usefulness to experimental psychologists was limited. However, a number of applications to experimental data have appeared more recently (e.g., Anderson, 1973;Anderson, Silverstein, Ritz, & Jones, 1977;Eich, 1982Eich, , 1985Heath, 1983; Humphreys, Pike, Bain, & Tehan, in press;Kawamoto & Anderson, 1985;Knapp & Anderson, 1984;Lewandowsky & Hockley, 1987; McDowd & Murdock, 1986;Murdock, 1982Murdock, , 1983 Nilsson, 1986), so the models can clearly make contact with experimental data.A further test is whether distributed memory models predict any novel or unexpected empirical phenomena. We argue that they do, and illustrate with a particular example. The prediction to be tested is that in a recognition learning situation, replacing items recognized on one trial should facilitate recognition of the repeated items (' 'targets") on the next trial. This facilitation is relative to a control condition in which there is no replacement.More specifically, in the experimental or replacement condition, one replaces all items recognized on the test phase of Trial n -1 with a new sample of items for the This work was supported by Natural Sciences and Engineering Research Council of Canada Grant APA 146 to Bennet B. Murdock, lr. We would like to thank our colleagues for many helpful comments on this research. Requests for reprints should be sent to Bennet Murdock. Department of Psychology, University of Toronto, Toronto, ON M5S IAI, Canada. 91 study phase of Trial n. In the control or nonreplacement condition, all old items from Trial n -1 are repeated in the study phase of Trial n. The prediction is that performa...

Output and retrieval interference in the missing-number task

“…Although the results from Bostrom and Bower, Rock (1957), and Murdock and Babick (1961) paralleled those of the present study, more recently Murdock and Lamon (1988) have presented evidence that made a different set of conclusions possible. In a recognition learning paradigm in which recognized words on one trial were replaced on the next trial, the repetition of unrecognized words resulted in greater recognition (relative to recognition of replacement words) than did a control condition in which no replacement took place.…”

Lure similarity affects visual episodic recognition: Detailed tests of a noisy exemplar model

“…In fitting our data, NEMO consistently chose a smaller standard deviation for the memorial noise of the more recent stimulus (see Table 2). Borrowing from the literature on global-matching models of recognition memory (Murdock & Kahana, 1993b;Murdock & Lamon, 1988), we also allow NEMO to differentially weight the contribution of s 1 and s 2 in the summed-similarity calculation (this is done via the model's parameter). This increased weighting of recently experienced items may be viewed either as a decay process or as a mechanism of strategic control.…”

Models for Recall and Recognition

“…TODAM, CHARM, and the Matrix model provide examples of models in which simply adding more copies of each trace to memory may not improve memory: Both the mean and standard deviation of the retrieved signal rise together in such a way that performance does not change (Murdock & Lamon 1988;Shiffrin et ai 1990). At least two approaches have been used to solve this problem.…”

“…At least two approaches have been used to solve this problem. Hintzman (1986Hintzman ( , 1988 and Murdock & Lamon (1988;see also Murdock 1989) have proposed a probabilistic encoding assumption: Each feature of an item is encoded (stored) with a probability p that rises with presentation time. If not stored it is given a neutral value (or in a variant discussed by Shiffrin et al 1990, replaced by random value).…”