Memory II

Three long-term recognition memory studies were carried out to explore the basic processes of recognition memory and the ability of current models to generalize to new settings. The experiments varied stimulus type (words and pictures), list lengths (12 and 24) and tasks given at testing (single item recognition, two-alternative forced choice and four-way classification). All experimental conditions giving probabilities of correct and error responses were predicted accurately by the REM model of Shiffrin & Steyvers (1997). In the simplest approach, REM was applied to every condition with the same decision threshold and with two of its three original parameter values, but with the third value augmented by another to account for better picture recognition. This ability to generalize and account for data from such a large number of conditions is especially surprising because REM is missing many components known to play an important role in recognition memory. The ability of such a minimal model to predict well suggests the few components it does include are fundamental and important enough to produce a good approximation to most results from recognition memory studies. These processes are: 1) incomplete or error-prone storage (or both); 2) activation of each trace by its similarity-determined match to the test probe, producing a likelihood ratio that that trace is OLD; 3) OLD/NEW recognition decisions based on familiarity defined as the average of the trace likelihood ratios.

Recent research by Mohammed, Meyer-Grant and Shiffrin showed that the very simple four-parameter REM model predicts the major findings of recognition memory using single-item and double-item testing without fitting parameters to the data. Our new research requires refining REM to incorporate many missing processes: 1) Context used as a first stage filter to determine what memory traces are activated. 2) Different forms of context, including one that changes between lists, particularly at the start of testing, and between experimental phases, and another that does not change. 3) Reconstruction of context that can occur during testing. 4) Choice of appropriate context to retrieve. 5) Strengthening of traces recalled during testing. 6) Storage of new traces during testing when there is no prior trace or when a prior trace is not recalled. 7) Changes in the decision criterion during testing so that the ratio of correct rejections to hits come closer to matching the experimental ratio of foils to targets. The new research had a first phase with ten lists of 20 pictures each followed by 10 tests of targets and 10 foils. A final test defined targets to be all items studied or tested in the first phase, all of which were tested mixed with an equal number of new foils. Targets in final testing were in random order, or order of lists, or the reverse, the latter two orders known to the participant.

Humans have a remarkable ability to focus on a specific part of their past and recall it in detail. Not only is one able to retrieve recent information, but also transport farther into the past and retrieve distant information with little interference. List-learning paradigms provide insight into these memory search processes, as one displays rich behaviors when searching within a list as well as when switching between lists to retrieve target information. However, existing computational models of memory search largely focus on capturing the behavior of either within-list or across-list recall. Here, we build a unified model of memory search by extending the Context Maintenance and Retrieval (CMR) Model. We propose that the starting context of each list serves as an entrypoint to a given list during memory search. One maintains access to the start of each list during study, and then at recall, jumps back to the start of the list to focus their retrieval on a specific list. We show that with this mechanism, the model is able to capture behavioral patterns of recall initiation, recall transition, and recall intrusion across multiple paradigms, including immediate and delayed free recall tasks (which place a memory demand on the most recent list), the externalized dual-list free recall task (which requires jumping back to a previous list), and the final free recall task (which requires switching memory search across multiple lists). We discuss the contributions of our model of within-list and across-list recall in relation to existing models of memory search.

Our everyday decisions are driven by a simultaneous search of our current environment and through our related memory bank. A simple example of this could be our search through a long restaurant menu guided by our recall of previous dining experiences. The influence of prior restaurant outings may aid in simplifying the menu, as you recall antipasto as an appetizer after a prior experience believing it to be a type of pasta. Furthermore, we do not recall every dining experience while looking through the menu, as experience at an American restaurant may not have relevant, overlapping features. The evidence accumulation literature embeds these constraints on search at both the feature level and the exemplar level. The Exemplar Based Random Walk (EBRW) model demonstrates the joint impact of selective attention and relevant recall of exemplars when making categorization decisions. However, this model includes a static version of attention, rather than having attention change dynamically as the subject learns. The goal of this work is to implement aspects of selective exemplar recall into current models of category learning, such as the Adaptive Attention Representation Model (AARM). Not only will we analyze the subset of exemplars used to make decisions, but how partially encoded traces influence the exemplar recalled. Incorporating EBRW’s exemplar sampling into AARM offers insight into decision-making and order effects.

The advantage for strongly encoded items over weakly encoded items is nearly identical in mixed-strength and pure-strength lists, in verbal episodic recognition, known as the null list-strength effect (Ratcliff et al., 1980). Early models produced this by differentiation in local-trace models and models assuming orthogonal item representations. We recently proposed an alternative, continuum-based account (compatible with most models), within attentional subsetting theory. This view explains the classic result by assuming that working representations of item vectors consist of a small number of item features subsetted from a high-dimensional feature-space. This sparseness approximates orthogonality. But when the dimensionality of the feature-space of the strong-items' features is small, cross-item similarity is high, resulting in a larger strength advantage in mixed than pure lists (as in the production effect). Interestingly, the theory also can produce the opposite: a greater advantage in pure than mixed lists. This occurs when the weaker condition (such as 500 ms display time of a word) has mainly shallow (compact-space) features stored and when those can be disregarded in pure-strong lists when deeper, sparse features can support recognition. We have replicated this inverted list-strength effect eight times (pre-registered). Quantitative fits to the data suggest that participants also adapt their response bias to list composition. Even massed repetition (3 x 500 ms displays) produces an inverted list-strength effect. One final experiment revealed a boundary condition: spaced repetitions, even with 500 ms study times, produced a null list-strength effect. Repeated, spaced presentations may render disregarding shallow features difficult, leaving only the approximate orthogonality of strongly encoded items that produces minimal similarity-based confusion across items.