Optimal Foraging in Semantic Memory

Abstract

When searching for items in memory, people explore internal representations in much the same way that animals forage in space. Results from a number of fields support this notion at a deeper level of evolutionary homology, with evidence that goal-directed cognition is an evolutionary descendent of animal foraging behavior (Hills, 2006). Is it possible then that humans forage in memory using similar search policies to the way that animals forage in space? To investigate this, we examine how people retrieve items from memory in the category fluency task: Participants were asked to retrieve as many types of animals from memory as they could in 3 minutes. Clusters or patches of these items, along with their semantic similarity and frequency, were found with an automatic Wikipedia corpus analysis using the BEAGLE semantic memory model (Jones & Mewhort, 2007), and via hand-coded category membership from Troyer et al. (1997). Participants did not seem to use static patch boundaries, such as 'pets', to search memory, but instead used fluid patch boundaries that were updated with each new item retrieved. We found that participants leave patches in memory when the marginal (i.e., current) rate of finding items is near the average rate for the entire task, as predicted by optimal foraging theory. Furthermore, participants appear to search within patches using item similarity, but decide where to "land" when moving between patches using item frequency.

The news is American but our memories are … Chinese?

“…Using a cognitive memory model, our results show that national news did favorably predict the order in which countries were recalled by individuals who were most likely to have read that (or a related) national news source. This suggests cognitive representations based not only on frequency of encounter in the environment (e.g., Anderson & Schooler, ) but also on the associative structure of items in relation to one another (e.g., Hills & Pachur, ; Hills et al., ). More important, these representations reflect their nation of origin.…”

“…To investigate the alignment between national news and national recall patterns, we used a variant of a prominent memory model (SAM; Raaijmakers & Shiffrin, ). SAM has been shown to predict how people recall animals (based on the co‐occurrence of animal names in a Wikipedia corpus; Hills, Jones, & Todd, ) and social associates (based on the structure of individual social networks; Hills & Pachur, ). Here, we use SAM to quantify the predictive power of each national news source with respect to the order in which countries were recalled.…”

“…Following previous work (e.g., Raaijmakers & Shiffrin, ; Hills et al., ), we used two cues representing global and local aspects of memory search. The global cue (representing the context of “countries”) activates each country in proportion to its overall frequency in the news: A ( Q , I ) = freq( I ) for each country, I , based on that country's frequency of occurrence for a given news source, with Q here simply representing the category “countries.” The local cue is the most recently recalled country (after the first recall) and activates other countries in proportion to their co‐occurrence with that country in the news: Here, the probe cue is the previously generated item, Q i = I t − 1 , and A ( Q , I ) = S ( I t − 1 , I ), with S ( I t − 1 , I ) equal to the co‐occurrence of the two countries for a given news source.…”

“…Long‐term memories reflect associations built in memory over repeated exposures to information. Recent investigations using cognitive memory models have shown that memories can be predicted using large text corpora, based on the principle that memory records associations between co‐occurring stimuli in the external environment (Hills, Jones, & Todd, ; Hills, Mata, Wilke, & Samanez‐Larkin, ). Importantly, however, memories are less influenced by massed exposure than by spaced exposure.…”

“…SAM, as proposed by Raaijmakers and Shiffrin (), uses sampling‐with‐replacement, whereas here we use sampling‐without‐replacement (see also Hills & Pachur, ; Hills et al., ). None of our conclusions hinge on this assumption.…”