A single-system model predicts recognition memory and repetition priming in amnesia

Abstract

We challenge the claim that there are distinct neural systems for explicit and implicit memory by demonstrating that a formal single-system model predicts the pattern of recognition memory (explicit) and repetition priming (implicit) in amnesia. In the current investigation, human participants with amnesia categorized pictures of objects at study and then, at test, identified fragmented versions of studied (old) and nonstudied (new) objects (providing a measure of priming), and made a recognition memory judgment (old vs new) for each object. Numerous results in the amnesic patients were predicted in advance by the single-system model, as follows: (1) deficits in recognition memory and priming were evident relative to a control group; (2) items judged as old were identified at greater levels of fragmentation than items judged new, regardless of whether the items were actually old or new; and (3) the magnitude of the priming effect (the identification advantage for old vs new items) overall was greater than that of items judged new. Model evidence measures also favored the single-system model over two formal multiple-systems models. The findings support the single-system model, which explains the pattern of recognition and priming in amnesia primarily as a reduction in the strength of a single dimension of memory strength, rather than a selective explicit memory system deficit.

Keywords: amnesia; computational model; long-term memory; memory systems; recognition memory; repetition priming.

Figure 1.

Example of a fragmented stimulus used in the identification portion of a CID-R trial at test. An object was initially presented at a highly fragmented level (level 1). Participants were instructed to try to identify the item at the most fragmented level they could. If the item could not be identified, a button press revealed a less fragmented version of the object (up to level 10).

Figure 2.

Model representations and Predictions 2 and 3. The top panels illustrate the relationship between the ID (identification level) and J_r variables in the models. The ellipses represent bivariate normal distributions of each class of item (old or new), cut horizontally and centered on a point that represents the mean J_r and ID for that class of item. Prediction 2 concerns whether ID levels are facilitated for items judged old within new and old items, that is, whether the mean ID of false alarms is less than that of correct rejections [CRs; i.e., CR − false alarm (FA)], and whether the mean ID of hits is less than of misses (i.e., MISS − HIT), where a correct rejection is a new judgment to a new item, a false alarm is an old judgment to a new item, a miss is a new judgment to an old item, and a hit is an old judgment to an old item. Prediction 3 concerns whether the priming effect overall (across all items) is greater than the priming effect for items judged new. Priming is calculated as mean ID(new items) − mean ID(old items); priming for items judged new is calculated as mean ID(CR) − mean ID(FA). The SS model predicts positive differences between ID(CR) − ID(MISS), ID(MISS) − ID(HIT), and priming − priming items judged new. The MS1 model predicts no differences. The MS2 model predicts positive differences when the explicit and implicit strengths of an item are positively correlated (i.e., w > 0), and predicts no differences when there is no correlation (i.e., w = 0).

Figure 3.

Recognition and priming task performance. a, Proportion of hit and false-alarm responses in the KOR and CON groups. b, Fragment identification performance according to whether the object during the test is actually new or old, or judged new or old. c, Fragment identification performance classified according to the recognition response [correct rejection (CR), miss, false alarm (FA), hit] in the KOR and CON groups. Bars indicate experimental data (error bars indicate 95% confidence intervals of the mean). Symbols indicate the expected result from each model when fit to data aggregated across individuals (a and b; because the data in these figures are derived from all of the participants) or the mean expected result from each model when fit to each individual's data (c; because the data in these figures are derived from the subset of participants with responses in all four recognition categories). In c, the letters represent the individuals in each group.

Figure 4.

Model prediction results. a, Recognition discrimination (P_r: proportion of hits − proportion of false alarms) and priming (i.e., fragment identification advantage for old objects) for the KOR and CON groups. Fluency effects (i.e., fragment identification advantage for objects judged old) across all items are also presented. Prediction 1 of the SS model is confirmed by lower recognition and priming in the KOR group than in the CON group. b, Differences in the ID level for items judged old versus judged new within new and old item types, and differences in the priming effect (overall) and the priming effect of items judged new. Predictions 2 and 3 of the SS model are confirmed in the KOR group. Bars indicate experimental data (error bars indicate 95% confidence intervals of the mean). Symbols indicate the expected result from each model when fit to data aggregated across individuals (a; because the data are derived from all of the participants) or the mean expected result from each model when fit to each individual's data (b; because the data are derived from the subset of participants with responses in all four recognition categories).

Figure 5.

Model selection results. Each bar represents the percentage of participants best fit by each model according to the AIC and the BIC in the CON and KOR groups. The SS model was the best-fitting model for the majority of participants, with the remainder being best fit by the MS1 model.

Figure 6.

Best-fitting models for each participant (according to the AIC; individual level fits). a, b, The best fitting models are plotted according to P_r and priming (mean identification new − mean identification old) performance (a) and the difference in ID levels for items judged old and new (i.e., fluency effects) within old and new items (b). It is evident that the participants in the KOR group who were best fit by the MS1 model tended to show priming (or recognition) in the near absence of recognition (or priming). The MS1 model can reproduce such a pattern because the μ_r|old and μ_p|old parameters can vary independently of one another. In the CON group, there were also participants who were best fit by the MS1 model even though they showed relatively large positive recognition and priming effects. These participants tended to show an absence of fluency effects (or even a negative fluency effect) within old or new items (b, right). Because f_p and f_r are uncorrelated in the MS1 model, it does not predict fluency effects within old/new items. Thus, the participants best fit by the MS1 model appeared to exhibit results that were consistent with its predictions. The letters A, B, and C above the points in the KOR group label patients who showed priming effects despite performing very close to chance in recognition.

Figure 7.

Performance of the KOR group patients A, B, and C (as labeled in Figs. 3c, 6). a–c, Recognition (a), priming (b), and differences (c) in ID levels for items judged new and old within old and new items (i.e., fluency effects), and differences in the priming effect (overall) and the priming effect of items judged new (Predictions 2 and 3 of the SS model). Bars denote data, and symbols indicate the expected result from each model when fit to the data from each individual. The dashed lines in a and b indicate the lower 95% confidence interval for the mean recognition and priming performance, respectively, in the CON group (Fig. 4).