The co-evolution of knowledge and event memory.

Abstract

We present a theoretical framework and a simplified simulation model for the co-evolution of knowledge and event memory, both termed SARKAE (Storing and Retrieving Knowledge and Events). Knowledge is formed through the accrual of individual events, a process that operates in tandem with the storage of individual event memories. In 2 studies, new knowledge about Chinese characters is trained over several weeks, different characters receiving differential training, followed by tests of episodic recognition memory, pseudo-lexical decision, and forced-choice perceptual identification. The large effects of training frequency in both studies demonstrated an important role of pure frequency in addition to differential context and differential similarity. The SARKAE theory provides a framework within which models for various tasks can be developed; we illustrate the way this could operate, and we make the verbal descriptions of the theory more precise with a simplified simulation model applied to the results.

A Dynamic Approach to Recognition Memory

“…Access to semantic memory depends not only on the physical features of the stimulus, such as the letters of a word, but can also affected by context, as has been found in studies of lexical decision and naming (Wagenmakers et al, 2004). Finally, although we do not model it here, just as semantic memory is used to determine the features encoded in an episodic trace, the episode augments the knowledge stored in semantic memory (Nelson & Shiffrin, 2013). This process of forming a memory trace that stores both content and context features is depicted schematically in Figure 2.1.…”

“…Conversely, given a sufficiently large representation, composite trace models can approximate the behavior of separate trace models; essentially, with a large enough composite vector, representations of individual memories become sparse enough to easily tell apart (Plate, 2003). There is also no reason to believe that composite-and separate-trace representations are mutually exclusive: Simultaneous storage of a sparse, distinct event trace along with addition to a growing composite trace underlies a variety of theories of the distinction between episodic and semantic memory (Nelson & Shiffrin, 2013), and is supported by differences in representation and function found between medial temporal lobes and neocortex (McClelland, McNaughton, & O'Reilly, 1995;Norman & O'Reilly, 2003;Kumaran & McClelland, 2012).…”

“…It is also straightforward to include pre-experimental memory traces in a separate-trace model such that context cues alone are no longer sufficient to isolate the memory traces relevant for the recognition task, an approach we will take in the model we present here. Finally, prior experience can influence how items are encoded (Nelson & Shiffrin, 2013), thus allowing for a certain type of longterm context noise. We defer fuller discussion of these issues until after we present our model.…”

“…Although history traces match in content features, they mismatch in context. Just as adding more list traces-which match in context, but not content-leads to a mirror effect, adding more history traces will do the same by increasing the odds that one of those traces was formed in a context similar to the experimental study context (higher false alarms to HF words) and introducing many poor matches that swamp the influence of a well-matching target trace (lower hit rate to HF words; Nelson & Shiffrin, 2013). Regardless of implementation, the contexts in which a word appears in prior life history affect episodic recognition: When normative word frequency is held constant but contextual diversity is varied, words that occur in more diverse contexts have a lower hit rate and higher false alarm rate than words that only occur in a few contexts (Steyvers & Malmberg, 2003).…”

Parallel interactive retrieval of item and associative information from event memory

“…Over time, the ability to store and retrieve information about the co-occurrence of events underlies the learning of a great deal of semantic information, such as word learning (co-occurrence of words and referents; Smith & Yu, 2008;Smith et al, 2014), causal learning (Zacks & Tversky, 2001), and learning whole object representations (Czerwinski et al, 1992;Gauthier & Tarr, 1997;Shiffrin & Lightfoot, 1997;Nelson & Shiffrin, 2013). In a broad sense, most memory begins as "associative" in that events and objects are collections of isolated entities.…”