Homogeneity of Memory Errors in Abstract Visual Pattern Recall

It is well established that human memory is not just an objective, unbiased storage system that records all incoming sensory information. Amongst other factors, it can be influenced by the direction of attention (e.g., Mulligan, 1998), by the amount of processing applied to stimuli (Craik & Lockhart, 1972), and by an individual’s familiarity with study material (Chase & Simon, 1973; Miller, 2003). The last of these factors is linked to the notion of a schema, where information that is aligned with knowledge and expectations can be easier to memorise (Alba & Hasher, 1983; Ghosh & Gilboa, 2014). Schemas have been hypothesised to encourage other mnemonic processes, for example by directing attention towards important aspects of stimuli (Alba & Hasher, 1983) or by enriching the representation and processing of stimuli (Miller, 2003).

In many circumstances, individuals use schematic knowledge to support memory, for example, by using a loved one’s name as a password or by using a meaningful date as a pin number. Empirical studies have shown that familiar material is easier to learn; for example, native words are easier to memorise than foreign words and nonwords (Hulme, Maughan, & Brown, 1991), native proverbs are easier to memorise than foreign proverbs (Poppenk, Kohler, & Moscovitch, 2010), and common visual scenes are easier to memorise than uncommon visual scenes (Badham, Hay, Foxon, Kaur, & Maylor, 2016). This also extends to individual differences in familiarity; for example, expert chess players are better able to visually memorise chess positions than novice players (Chase & Simon, 1973), and individuals more experienced with a given topic have been shown to be more able to memorise text based on that topic than less experienced individuals (Arbuckle, Vanderleck, Harsany, & Lapidus, 1990; Miller, 2003).

Early research by Bartlett (1932) postulated memory as a reconstructive process where schematic information was used to fill in gaps during recall. His work showed that after studying an unfamiliar folk tale, participants’ recall attempts often included simplification and distortions which aligned the details of the story with their own knowledge and culture. Modern research has built upon the work of Bartlett using his serial reproduction technique where the recalled information from one participant is presented as the study material for the next participant and so on. Over generations of serial reproduction, cultural studies have shown systematic shifts in the content of information, indicating that participants are imparting their own meaning onto study materials (Bangerter, 2000; Mesoudi & Whiten, 2004). Studies of language acquisition have explored serial reproduction with novel material in order to understand inductive biases in language learning. A key feature of such work is the fact that later generations of serial learners show more regularity in the material they report, demonstrating that individuals are imparting structure upon the information that they are learning (Reali & Griffiths, 2009; Smith & Wonnacott, 2010). This indicates that schema-like organisational processes are present in novel language learning, and that such processes are not unique to paradigms where schemas/meanings are concrete and explicitly identifiable.

Serial reproduction has also been investigated with more abstract stimuli. Ravignani, Delgado, and Kirby (2016) asked participants to learn random rhythmic patterns and their recall data were presented to later participants as study material and so on. Across generations of learners, patterns became more structured and easier to learn. Kalish, Griffiths, and Lewandowsky (2007) trained participants to learn functions mapping a stimulus magnitude x to a response magnitude y. Participants were trained on linear and nonlinear functions mapping x to y as well as random mappings. Across generations of learners, linear functions emerged in participants’ responses regardless of the function learned by the first generation (see also, Ferdinand & Zuidema, 2008). This indicated that a bias systematically distorted participants’ memory towards a simple relationship between x and y. Another study showed that, across generations of serial reproduction, participants tended to learn novel categorical information with a bias towards simpler organisational structures (Griffiths, Christian, & Kalish, 2008). The influence of an inductive bias has also been experimentally manipulated. Xu and Griffiths (2010) trained participants to learn sizes of imaginary fish for a classification task. Following this, participants completed a perceptual task where images of the fish were very briefly presented on the screen and participants were required to reconstruct their size. The reconstructions were used for later generations of participants and, across generations, the reconstructed sizes converged towards whatever size that group of participants had learned in the initial classification task.

Whilst serial reproduction tasks have been utilised to identify and classify inductive biases, there has been less emphasis on investigating the mnemonic properties of reproduced stimuli. Given that serial reproduction biases often converge towards prior knowledge and experience, and that prior knowledge and experience is known to support memory reconstruction, the current study aimed to establish if recall attempts for a given random stimulus (reconstructions) can be more memorable than the original stimulus. Ravignani et al. (2016), discussed above, established that recall attempts of random rhythmic patterns were easier to learn than the original patterns and other studies have shown improved learning in later generations of language learners (e.g., Kirby, Cornish & Smith, 2008; see Brighton Kirby & Smith, 2005, for a review). These previous studies have used multiple generations of learners, so it may be the case that a survival-of-the-fittest mechanism biases learning where easy to learn material is successfully transferred between learners but material more difficult to learn is lost. Crucially, this effect could be driven by biases in encoding, not retrieval. The current study focusses on retrieval by observing mnemonic effects across just one generation of learners, and by utilising a memory test where retrieval distortions can be measured.

Furthermore, to maximise the amount of structure participants could impart to stimuli during reproduction, the initial study material in the current study was entirely random, this is aligned with existing serial reproduction studies using novel stimuli. However, by utilising a paradigm based on the Visual Patterns Test (Della Sala, Gray, Baddeley, Allamano, & Wilson, 1999), our stimuli were richer than the novel stimuli in previous studies of serial reproduction and could be incorrectly recalled in many ways, affording greater opportunity for observation of retrieval biases. We also included a difficulty manipulation (study-test delay) in order to establish if mnemonic effects of reproduction are greater when recall is harder.

Section 1: Mnemonic Tests

Section 2: Homogeneity Tests

In order to quantitatively assess recall biases, analyses were conducted to assess the similarity of participants’ responses to 1) the responses of other participants and 2) responses to other trials within each participant’s own data. For every trial, the recall data for a “comparison” trial was compared to all “alternative” trials where comparison and alternative trials were never based on the same study patterns. Similarity was therefore calculated comparing the comparison trial to 1) all alternative recall data from other participants only or 2) all alternative recall data from that same participant only. As a baseline, the same comparisons were made but instead of comparing the comparison trial to alternative recall data it was compared to the alternative original studied patterns. Therefore, the only differences between the initial similarity measures (1 and 2) and their corresponding baseline measures were the errors made during recall. This was done to establish if the similarity measures were above baseline, which would indicate homogeneity of recall errors in the data.

Similarity was calculated following Nosofsky (1984). Each element of the recall data from the comparison trial was coded as a vector (t) of 16 ones and zeros representing the recall of black or white sections of a given 4 × 4 pattern. A corresponding vector (x) was created for each alternative trial to which the comparison trial was compared. The similarity between these two vectors was computed using the following equation:

where j cycles through each element of the vectors. For the current purposes, r was set to two to give an Euclidian distance metric (Nosofsky, 1984) and c was set to one. The average of these similarity measures was computed for the measures outlined above for 1) all comparisons to other participants 2) all comparisons to other trials from the same participant. The two baseline average similarities corresponding to (1) and (2) were computed based on the exact same trials but using the original study patterns instead of recall data in alternative trials. All of these four measures were computed separately using just instant recall trials or just delayed recall trials resulting in a 2 × 2 × 2 design (see below).

As the above measures were computed for random study patterns only, Experiments 1 and 2 were comparable and were entered into the same statistical analysis. A 2 (comparison group; other participants, same participant) × 2 (comparison type; recall data, baseline/studied data) × 2 (memory test; instant, delayed) repeated measures ANOVA was computed on the mean summed similarity measures for each participant (see Figure 2, for descriptive statistics).¹ A main effect of comparison group indicated that similarity was higher between data from the same participant than between a given participant’s data and other participants’ data, F(1, 67) = 14.17, MSE = 1.82 × 10^-4, p < .001, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = .18. A main effect of comparison type indicated homogeneity across recall responses—similarity was higher for comparisons between recall data than for comparisons of recall data to baseline/studied data, F(1, 67) = 11.48, MSE = 9.45 × 10^-5, p = .001, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = .15. There was no main effect of memory test, F(1, 67) = 2.76, MSE = 2.83 × 10^-4, p = .102, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = .04. Interestingly, an interaction between comparison group and comparison type revealed larger differences between recall data and baseline/studied data similarity measures for data from the same participant than for data from other participants, F(1, 67) = 12.85, MSE = 8.63 × 10^-5, p = .001, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = .16. This indicated homogeneity within the participants’ responses (similarity in excess of baseline similarity) was greater than homogeneity across different participants’ responses. Generally, homogeneity of responses only occurred for data based on delayed tests for comparisons within a single participant’s data (see Figure 2): There was an interaction between comparison group and memory test, F(1, 67) = 3.99, MSE = 2.38 × 10^-4, p = .0497, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = .06, between comparison type and memory test, F(1, 67) = 10.33, MSE = 9.24 × 10^-5, p = .002, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = .13, and a three-way interaction between all factors, F(1, 67) = 10.45, MSE = 8.23 × 10^-5, p = .002, ${\mathrm{\eta }}_{\mathrm{p}}^2$ = .14. This pattern was confirmed by follow up paired t-tests across the comparison-type factor (comparing the mean similarity measures based on recall data as alternative trials to similarity measures based on baseline/studied data as alternative trials). For the following four t-tests, a Bonferroni correction was utilised with alpha set to .0125 for determining significance. For similarity to other participants, neither instant, t < 1, or delayed, t(67) = 1.07, p = .289, trials showed similarity in excess of baseline similarity. For similarity to other trials from the same participant, instant tests showed no similarity in excess of the baseline similarity, t(67) = 1.04, p = .314, but for delayed tests, there was significantly more similarity between participant’s own recall data than between recall data and baseline similarity, t(67) = 3.42, p = .001.

The homogeneity of errors observed could be simply due to misremembering as opposed to there being an innate structure to memory errors as one possible reason for the homogeneity of errors observed could be intrusions. Although we never calculated similarity measures between two sets of data based on the same study pattern, if a participant’s recall was an intrusion then this could enhance similarity if that intrusion was compared to the pattern on which the intrusion was based. For example, if a comparison trial was “Pattern 1” then this would be compared to all other alternative patterns (Patterns 2–24) to establish mean similarity. Therefore, if the participant failed to recall Pattern 3 but instead recalled an intrusion of Pattern 1, then this intrusion would have high similarity to Pattern 1 and would increase the overall mean similarity based on Patterns 2–22. To account for this, a different baseline measure was computed.

In the earlier analyses, comparison trials were based on recall data. The homogeneity of errors observed could have been due to an innate bias in the errors produced or due to intrusions. If it was due to intrusions, we would see the same enhanced similarity if comparison trials were recall data or if comparison trials were original studied data. This is because an intrusion would show high similarity to a previous trial regardless of whether the similarity measurement was based on recall data or the pattern that was originally studied as both would resemble the intrusion. Therefore, by utilising comparison trials using original studied data, a new baseline measure is formed and homogeneity above this indicates homogeneity unlikely to be due to intrusions. The measures of similarity were recomputed using original studied data as comparison trials for delayed recall only and for comparisons within the same participant only (i.e., for the only data that showed homogeneity—see Figure 2). With the new comparison trials, there was no significant difference between similarity based on alternative recall data versus similarity based on alternative original study patterns, t < 1 (M = 0.067, SD = 0.011; M = 0.066, SD = 0.010, respectively). Therefore, there was no indication that similarity was enhanced by intrusions. Furthermore, enhanced similarity across recall data was still present for delayed trials within a participant’s own responses, when compared to the new baseline measure that accounted for intrusions, t(67) = 3.13, p = .003. This indicates that the homogeneity observed in the earlier analyses was not due to intrusions but was due to an innate bias in the errors made by each participant.

General Discussion

Across two experiments testing memory for visual matrix patterns, participants recalled patterns that were either random or were previous participants’ attempts to recall random patterns. It was initially hypothesised that previous participants’ recall attempts would be more memorable than random patterns because their recall errors may be influenced by reconstruction biases in individuals, reflecting what people are better able to learn. Experiment 1 showed some evidence of this with significantly better recall for study patterns based on previous participants’ recall than for random patterns, but this only occurred with items analysis and not for participants analysis and the effect did not replicate in Experiment 2, which had a larger data set. Despite this, there was evidence of homogeneity in the errors made by participants; when they made recall errors, the errors were similar across different trials when measured within participants. Furthermore, homogeneity of errors occurred within but not between participants and it could not be explained by intrusions. Overall, it appears that participants do show a bias in the way they recall visual matrix patterns. However, the bias may be unique to each individual and may lead to the production of patterns that are not particularly memorable to other individuals.

One aspect of the current study that differs from previous research based on learning material generated by other participants is the use of visual patterns as stimuli. The majority of research focussing on iterated learning over generations of participants is focussed on investigating language development (e.g., Kirby, Griffiths, & Smith, 2014). This is of interest to researchers because the acquisition of a language is an iterative process outside the laboratory (Kirby et al., 2008) with languages being transmitted throughout generations of individuals. Therefore, the abstract visual stimuli utilised in the current study may not have evoked the utilisation of mnemonic mechanisms hypothesised to have evolved specifically for the acquisition of language (cf. Pinker, 1994).

In contrast to the above, some previous studies do show improved learning in later generations using non-verbal material (e.g., rhythms, Ravignani et al., 2016; drawings, Tamariz, & Kirby, 2015) but mnemonic benefits seem to take multiple generations to manifest. Tamariz and Kirby argued that memory processes result in material becoming more compressible over iterations of learners. For example, in their study abstract drawings began to resemble letters across generations, allowing individuals to reproduce the shapes using their knowledge of the symbols. We have every reason to expect that a multiple-generation version of the current study would yield similar results, especially given that visual matrix patterns have been shown to be more memorable when they could be represented verbally (Brown et al., 2006). Therefore, we believe that mnemonic effects may take multiple generations to manifest with visual stimuli, where random memory errors eventually converge on schema-consistent patterns that resist degradation across generations. Although the use of a single generation was a potential limitation of the current study, it did allow us to utilize a much larger set of stimuli than those usually employed in multiple generation studies and to explore a generation of iterative learning in more detail, such as the comparison between a single individual’s responses.

Given that there was similarity in the errors made within an individual’s set of responses in the current data, it is highly likely that they would show higher accuracy in a memory test based on their own responses from earlier trials than they would show in a memory test based on random data. This is because in the absence of accurate recall, retrieval attempts based on guessing would match what was actually studied for a memory test that included guessing from earlier trials. That is, there appears to be consistency in guessing in memory tests for simple abstract visual stimuli. This would not be easy to test for as it would be difficult to dissociate from the memory benefits provided by repeated exposure to similar stimuli (cf. Hintzman & Block, 1971). Despite this, the current data are partly aligned with the generation effect, where material that is produced by an individual is easier to retrieve later by that same individual than material that is passively studied (e.g., Bertsch, Pesta, Wiscott, & McDaniel, 2007; Slamecka & Graf, 1978). A typical example of the generation effect involves generating a word from a clue; for example, finding a synonym of the word rapid beginning with f to get the target word fast. In later memory tests such target words are recalled better in generation conditions than in control conditions where a different participant simply reads the word pair rapid-fast (Slamecka & Graf, 1978). One mechanism by which the generation effect is proposed to operate is through transfer-appropriate processing; where the act of generation resembles processes that take place during retrieval, aiding memory (Bertsch et al., 2007). It may be the case that in the absence of memory for simple abstract visual stimuli, participants guess the pattern at retrieval. Then, when they forget parts of a pattern from a later trial, they are inclined to guess in the same way. This could explain why we saw more homogeneity within participants than between participants as the guessing would be highly dependent on the initial trials which were randomised for different participants.

Even though our data showed significantly more homogeneity of errors within participants than between participants, at least one iterated learning study has found largely equivalent inductive biases when comparing within- and between- participants designs (Griffiths et al., 2008). Given this discrepancy and the ideas expressed above, we tentatively propose two mechanisms that could bias the way information is stored and retrieved across generations of learners: 1) a survival-of-the-fittest mechanism where across multiple generations of learners, information that is easier to encode is more likely to be successfully transmitted to the next generation and 2), reconstruction biases with systematic influences on guessing in the absence of memory. This second mechanism could operate similarly to the generation effect as described above or could simply be that some stimuli are easier to guess than others. For example, as summarised in the introduction, Kalish et al. (2007) found that across generations of learners different x-y functions ended up as positive linear functions—it is established that simple linear functions are easier to learn than other types of function (Brehmer, 1971), possibly because this is the relation participants are most likely to guess (Naylor & Clark, 1968). It is therefore important for future researchers aiming to understand inductive biases to dissociate 1) effects based on what is easier to encode from 2) influences on guessing in the absence of memory.