{"title": "A simple model of recognition and recall memory", "book": "Advances in Neural Information Processing Systems", "page_first": 293, "page_last": 301, "abstract": "We show that several striking differences in memory performance between recognition and recall tasks are explained by an ecological bias endemic in classic memory experiments - that such experiments universally involve more stimuli than retrieval cues. We show that while it is sensible to think of recall as simply retrieving items when probed with a cue - typically the item list itself -  it is better to think of recognition as retrieving cues when probed with items. To test this theory, by manipulating the number of items and cues in a memory experiment, we show a crossover effect in memory performance within subjects such that recognition performance is superior to recall performance when the number of items is greater than the number of cues and recall performance is better than recognition when the converse holds. We build a simple computational model around this theory, using sampling to approximate an ideal Bayesian observer encoding and retrieving situational co-occurrence frequencies of stimuli and retrieval cues. This model robustly reproduces a number of dissociations in recognition and recall previously used to argue for dual-process accounts of declarative memory.", "full_text": "A simple model of recognition and recall memory\n\nNisheeth Srivastava\n\nComputer Science, IIT Kanpur\n\nKanpur, UP 208016\n\nnsrivast@cse.iitk.ac.in\n\nEdward Vul\n\nDept of Psychology, UCSD\n\n9500 Gilman Drive La Jolla CA 92093\n\nevul@ucsd.edu\n\nAbstract\n\nWe show that several striking differences in memory performance between recogni-\ntion and recall tasks are explained by an ecological bias endemic in classic memory\nexperiments - that such experiments universally involve more stimuli than retrieval\ncues. We show that while it is sensible to think of recall as simply retrieving\nitems when probed with a cue - typically the item list itself - it is better to think\nof recognition as retrieving cues when probed with items. To test this theory, by\nmanipulating the number of items and cues in a memory experiment, we show\na crossover effect in memory performance within subjects such that recognition\nperformance is superior to recall performance when the number of items is greater\nthan the number of cues and recall performance is better than recognition when\nthe converse holds. We build a simple computational model around this theory,\nusing sampling to approximate an ideal Bayesian observer encoding and retrieving\nsituational co-occurrence frequencies of stimuli and retrieval cues. This model\nrobustly reproduces a number of dissociations in recognition and recall previously\nused to argue for dual-process accounts of declarative memory.\n\n1\n\nIntroduction\n\nOver half a century, differences in memory performance in recognition and recall-based experiments\nhave been a prominent nexus of controversy and confusion. There is broad agreement among\nmemory researchers, following Mandler\u2019s in\ufb02uential lead, that there are at least two different types\nof memory activities - recollection, wherein we simply remember something we want to remember,\nand familiarity, wherein we remember having seen something before, but nothing more beyond\nit [8]. Recall-based experiments are obvious representatives of recollection. Mandler suggested that\nrecognition was a good example of familiarity activity.\nDual-process accounts of memory question Mandler\u2019s premise that recognition is exclusively a\nfamiliarity operation. They argue, phenomenologically, that recognition could also succeed successful\nrecollection, making the process a dual composition of recollection and familiarity [20]. Experimental\nprocedures and analysis methods have been designed to test for the relative presence of both processes\nin recognition experiments, with variable success. These endeavors contrast with strength-based\nsingle-process models of memory that treat recognition as the retrieval of a weak trace of item\nmemory, and recall as retrieval of a stronger trace of the same item [19].\nThe single/dual process dispute also spills over into the computational modeling of memory. Gillund\nand Shiffrin\u2019s in\ufb02uential SAM model is a single-process account of both recognition and recall [4]. In\nSAM and other strength-based models of declarative memory, recognition is modeled as item-relevant\nassociative activation of memory breaching a threshold, while recall is modeled as sampling items\nfrom memory using the relative magnitudes of these associative activations. In contrast, McClelland\u2019s\nequally in\ufb02uential CLS model is explicitly a dual-process model, where a fast learning hippocampal\ncomponent primarily responsible for recollection sits atop a slow learning neocortical component\nresponsible for familiarity [9]. Wixted\u2019s signal detection model tries to bridge the gap between\n\n31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.\n\n\fthese accounts by allowing dual process contributions to combine additively into a unidimensional\nstrength variable [19]. While such pragmatic syntheses are useful, the \ufb01eld is still looking for a more\nsatisfactory theoretical uni\ufb01cation.\nThe depth of the difference between the postulated dual processes of recollection and familiarity\ndepends inevitably on the strength of the quantitative and qualitative dissociations that previous\nresearch has documented in memory tasks, prominent among which are recognition and recall.\nMandler, for instance, postulated a one-to-one mapping between recognition and familarity on one\nhand and recall and recollection on the other [8], although other authors hold more nuanced views [20].\nNotwithstanding such differences of opinion, the road to discovering useful single-process accounts\nof declarative memory has to go through explaining the multiple performance dissociations between\nrecognition and recall memory tasks. To the extent that single process accounts of both tasks can\nexplain such dissociations, differences between recollection and familarity will not seem nearly as\nfundamental.\nImproved strength-based models have competently modeled a large array of recognition-recall\ndissociations [13], but fail, or have to make intricate assumptions, in the face of others [20]. More\nimportantly, the SAM model and its descendants are not purely single-process models. They model\nrecognition as a threshold event and recall as a sampling event, with the uni\ufb01cation coming from\nthe fact that both events occur using the same information base of associative activation magnitude.\nWe present a much simpler single process model that capably reproduces many critical qualitative\nrecognition-recall dissociations. In the process, we rationalize the erstwhile abstract associative\nactivation of strength-based memory models as statistically ef\ufb01cient monitoring of environmental\nco-occurrence frequencies. Finally, we show using simulations and a behavioral experiment, that the\nlarge differences between recognition and recall in the literature can be explained by the responses of\nan approximately Bayesian observer tracking these frequencies to two different questions.\n\n2 Model\n\nWe use a very simple model, speci\ufb01ed completely by heavily stylized encoding and retrieval processes.\nThe encoding component of our model simply learns the relative frequencies with which speci\ufb01c\nconjunctions of objects are attended to in the world. We consider objects x of only two types: items\nxi and lists xl. We model each timestep as as a Bernoulli trial between the propensity to attend to\nany of the set of items or to the item-list itself, with a uniform prior probability of sampling any\nof the objects. Observers update the probability of co-occurrence, de\ufb01ned in our case rigidly as\n1-back occurrence, inductively as the items on the list are presented. We model this as the observer\u2019s\nsequential Bayesian updates of the probability p(x), stored at every time step as a discrete memory\nengram m.\nThus, in this encoding model, information about the displayed list of items is available in distributed\nform in memory as p(xi, xl|m), with each engram m storing one instance of co-occurrence. The true\njoint distribution of observed items,to the extent that it is encoded within the set of all task-relevant\nmemory engrams M is then expressible as a simple probabilistic marginalization,\n\np(xi, xl) =\n\np(xi, xl|m)p(m),\n\n(1)\n\nwhere we assume that p(m) is \ufb02at over M, i.e. we assume that within the set of memory engrams\nrelevant for the retrieval cue, memory access is random.\nOur retrieval model is approximately Bayesian. It assumes that people sample a small subset of all\nrelevant engrams M(cid:48) \u2282 M when making memory judgments. Thus, the joint distribution accessible\nto the observer during retrieval becomes a function of the set of engrams actually retrieved,\n\npMk (xi, xl) =\n\np(xi, xl|m)p(m),\n\n(2)\n\nwhere Mk denotes the set of \ufb01rst k engrams retrieved.\nFollowing a common approach to sampling termination in strength-based sequential sampling memory\nmodels, we use a novelty threshold that allows the memory retrieval process to self-terminate when\nincoming engrams no longer convey signi\ufb01cantly novel information [4, 13]. We treat the arrival of the\n\n(cid:88)\n\nm\u2208M\n\n(cid:88)\n\nm\u2208Mk\n\n2\n\n\fFigure 1: Illustrating the ecological difference in retrieval during recognition and recall memory\nexperiments. We model recall retrieval as a probabilistic query about items conditioned on the item\nlist and recognition retrieval as a probabilistic query about the item list conditioned on the item\npresented during retrieval. Since there are almost always more items than lists in classic memory\nexperiments, the second conditional distribution tends to be formed on a smaller discrete support set\nthan the former.\n\nkth successive engram into working memory as a signal for the observer to probabilistically sample\nfrom pMk. The stopping rule for memory retrieval in our model is for n consecutive identical samples\nbeing drawn in succession during this internal sampling, n remaining a free parameter in the model.\nThis rule is designed to capture the fact that memory search is frugal and self-terminating [15]. The\nsample drawn at the instant the novelty threshold is breached is overtly recalled. Since this sample is\ndrawn from a distribution constructed by approximately reconstructing the true encoded distribution\nof situational co-occurrences, the retrieval model is approximately Bayesian. Finally, since our\nencoding model ensures that the observer knows the joint distribution of event co-occurrences, which\ncontains all the information needed to compute marginals and conditionals also, we further assume\nthat these derivative distributions can also be sampled, using the same retrieval model, when required.\nWe show in this paper that this simple memory model yields both recognition and recall behavior.\nThe difference between recognition and recall is simply that these two retrieval modalities ask two\ndifferent questions of the same base of encoded memory - the joint distribution p(xi, xl). We illustrate\nthis difference in Figure 1. During recall-based retrieval, experimenters ask participants to remember\nall the items that were on a previously studied list. In this case, the probabilistic question being asked\nis \u2019given xl, \ufb01nd xi\u2019, which our model would answer by sampling p(xi|xl). In item-recognition\nexperiments, experimenters ask participants to determine whether each of several items was on a\npreviously shown list or not. We assert that in this case the probabilistic question being asked is\n\u2019given xi. \ufb01nd xl\u2019, which our model would answer by sampling p(xl|xi).\nOur operationalization of recognition as a question about the list rather than the item runs contrary to\nprevious formalizations, which have tended to model it as the associative activation engendered in\nthe brain by observing a previously seen stimulus - models of recognition memory assume that the\nactivation for previously seen stimuli is greater, for all sorts of reasons. In contrast, recall is modeled\nin classical memory accounts much the same way as in ours - as a conditional activation of items\nassociated with retrieval cues, including both the item list and temporally contiguous items. Our\napproach assumes that the same mechanism of conditional activation occurs in recognition as well -\nthe difference is that we condition on the item itself.\n\n3\n\nRecallRecognitionEncodingEncodingRetrievalRetrievalABCABCAp(x| )p( |A)Sample fromSample fromXABCp(x| )p( |x)\f3 Basic prediction: fast recognition and slow recall\n\nThe sample-based threshold used to terminate memory retrieval in our model \u0001 does not depend on\nthe size of the support of the probability distribution being sampled from. This immediately implies\nthat, for the same threshold sample value, the model will take longer to approach it when sampling\nfrom a distribution with larger support than when sampling from distributions with smaller support.\nIn classical memory experiments, observers are typically asked to memorize multiple items associated\nwith one, or a few, lists. Thus, there is an ecological bias built into classic memory experiments such\nthat |items| (cid:29) |lists|. Making this assumption immediately rationalizes the apparent difference in\nspeed and effort between recognition and recall in our model. Because the recognition task samples\np(list|item), its sample complexity is lower than recall, which involves sampling p(item|list) from\nmemory.\nTo verify this numerically, starting from identical memory encodings in both cases, we ran 1000\nsimulations of recognition and recall respectively using our retrieval model, using a \ufb01xed n = 5 1.\nThe results, measured in terms of the number of retrieval samples k drawn before termination in\neach of the 1000 trials, are shown in the left panel of Figure 2. The sample complexity of recall is\nevidently higher than for recognition2. Thus, we suggest that the fundamental difference between\nrecognition and recall - that recognition is easier and recall is harder - is explicable simply by virtue\nof the ecological bias of memory experiments that use fewer cues than stimuli.\nThe difference in speed between recollection and familiarity processes, as measured in recall and\nrecognition experiments, has been one of the fundamental motivations for proposing that two memory\nprocesses are involved in declarative memory. Dual-process accounts have invoked priority arguments\ninstead, e.g.\nthat information has to pass through semantic memory, which is responsible for\nrecognition, before accessing episodic memory which is responsible for recall [17].Single process\naccounts following in the lineage of SAM [4] have explained the difference by arguing that recognition\ninvolves a single comparison of activation values to a threshold, whereas recall involves competition\nbetween multiple activations for sampling. Our model rationalizes this distinction made in SAM-style\nsequential sampling models by arguing that recognition memory retrieval is identical to recall memory\nretrieval; only the support of the distribution from which the memory trace is to be probabilistically\nretrieved changes. Thus, instead of using a race to threshold for recognition and a sampling process\nin recall, this model uses self-terminating sampling in both cases, explaining the main difference\nbetween the two tasks - easy recognition and hard recall - as a function of typical ecological parameter\nchoices. This observation also explains the relative indifference of recognition tasks to divided\nattention conditions, in contrast with recall which is heavily affected [2]. Because of the lower sample\ncomplexity of recognition, fewer useful samples are needed to arrive at the correct conclusion.\n\n4 An empirical test\n\nThe explanation our model offers is simple, but untested. To directly test it, we constructed a simple\nbehavioral experiment, where we would manipulate the number of items and cues keeping the total\nnumber of presentations constant, and see how this affected memory performance in both recognition\nand recall retrieval modalities. Our model predicts that memory performance dif\ufb01culty scales up with\nthe size of the support of the conditional probability distribution relevant to the retrieval modality.\nThus recall, which samples from p(item|list), should become easier as the number of items to recall\nper cue reduces. Similarly recognition, which samples from p(listlitem), should become harder as\nthe number of cues per item increases. Because classic memory experiments have tended to use more\nitems than cues (lists), our model predicts that such experiments would consistently \ufb01nd recognition\nto be easier than recall. By inverting this pattern, having more cues than items, for instance, we would\nexpect to see the opposite pattern hold. We tested for this performance crossover using the following\nexperiment.\n\n1Our results are relatively independent of the choice of n, since for any value of n, recognition stays easier\n\nthan recall so long as the cue-item fan out remains large and vice versa.\n\n2Recall trials that timed out by not returning a sample beyond the maximum time limit (100 samples) are not\nplotted. These corresponded to 55% of the trials, resulting in a recall hit rate of 45%. In contrast, the average\nrecognition hit rate was 82% for this simulation.\n\n4\n\n\fFigure 2: (Left) Simulation results show easier recognition and harder recall given typical ecological\nchoices for stimuli and cue set sizes. (Right) Results from experiment manipulating the stimuli and\ncue set size ratio. By manipulating the number of stimuli and cues, we predicted that we would\nbe able to make recall harder than recognition for experiment participants. The results support our\nprediction unambiguously.Error bars show s.e.m.\n\nWe used a 2\u00d72 within subject factorial design for this experiment, testing for the effect of the retrieval\nmode - recognition/recall and either a stimulus heavy, or cue heavy selection of task materials. In\naddition, we ran two conditions between subjects, using different parameterization of the stimuli/cue\nratios. In the stimulus heavy condition, for instance, participants were exposed to 5 stimuli associated\nwith 3 cues, while for the cue heavy condition, they saw 3 stimuli associated with 5 cues. The semantic\nidentity of the stimuli and cue sets were varied across all four conditions randomly, and the order of\npresentation of conditions to participants was counterbalanced. All participants worked on all four\nof the memory tasks, with interference avoided with the use of semantically distinct category pairs\nacross the four conditions. Speci\ufb01cally, we used number-letter, vegetable-occupation, fruit-adjective\nand animal-place category pairs for the four conditions. Within each category, stimuli/cues for a\nparticular presentation were sampled from a 16 item master list, such that a stimulus could not occur\ntwice in conjunction with the same cue, but could occur in conjunction with multiple cues.\n120 undergraduates participated in the experiment for course credit. Voluntary consent was obtained\nfrom all participants, and the experimental protocol was approved by an institutional IRB. We told\nexperiment participants that they would be participating in a memory experiment, and their goal was\nto remember as many of the items we showed them as possible. We also told them that the experiment\nwould have four parts, and that once they started working on a part, there would be no opportunity to\ntake a break until it ended. 80 participants performed the experiment with 3/5 and 5/3 stimulus-to-cue\nmappings, 40 did it with 2/7 and 7/2 stimulus-to-cue mappings. Note that in all cases, participants\nsaw approximately the same number of total stimulus-cue bindings (3x5 = 15 or 2x7 = 14), thus\nundergoing equivalent cognitive load during encoding.\nStimuli and cues were presented onscreen, with each pair appearing on the screen for 3 seconds,\nfollowed by an ITI of equal duration. To prevent mnemonic strategy use at the time of encoding, the\nhorizontal orientation of the stimulus-cue pair was randomly selected on each trial, and participants\nwere not told beforehand which item category would be the cue; they could only discover this at\nthe time of retrieval3. Participants were permitted to begin retrieval at their own discretion once\nthe encoding segment of the trial had concluded within each condition. All participants chose to\ncommence retrieval without delay. Participants were also permitted to take breaks of between 2-5\nminutes between working on the different conditions, with several choosing to do so.\nOnce participants had seen all item-pairs for one of the conditions, the experiment prompted them to,\nwhen ready, click on a button to proceed to the testing phase. In the recall condition, they saw a text\nbox and a sentence asking them to recall all the items that occurred alongside item X, where X was\nrandomly chosen from the set of possible cues for that condition; they responded by typing in the\nwords they remembered. For recognition, participants saw a sentence asking them to identify if X had\noccurred alongside Y, where Y was randomly chosen from the set of possible cues for that condition.\n\n3An active weblink to the actual experiment is available online at [anonymized weblink].\n\n5\n\n{2,7}{3,5}{4,4}{5,3}{7,2}Condition00.511.522.5d'RecognitionRecall0501000100200300400Recognition0501000100200300400TrialsSample countRecall\fAfter each forced yes/no response, a new X was shown. Half the X\u2019s shown in the recognition test\nwere \u2019lures\u2019 , they had not been originally displayed alongside Y.\nMemory performance was measured using d\u2019, which is simply the difference between the z-normed\nhit rate and false alarm rate, as is conventional in recognition experiments. d\u2019 is generally not used\nto measure recall performance, since the number of true negatives is unde\ufb01ned in classic recall\nexperiments, which leaves the false alarm rate unde\ufb01ned as well. In our setup, the number of true\nnegatives is obviously the number of stimuli the participant saw that were not on the speci\ufb01c list\nbeing probed, which is what we used to calculate d-prime for recall as well.\nThe right panel in Figure 2 illustrates the results of our experiment. The predicted crossover is\nunambiguously observed. Further, changes in memory performance across the stimulus-cue set size\nmanipulation is symmetric across recognition and recall. This is precisely what we\u2019d expect if set\nsize dependence was symmetrically affecting memory performance across both tasks as occurs in\nour model. While not wishing to read too much into the symmetry of the quantitative result, we note\nthat such symmetry under a simple manipulation of the retrieval conditions appears to suggest that\nthe manipulation does in fact affect memory performance very strongly. Overall, the data strongly\nsupports our thesis - that quantitative differences in memory performance in recognition and recall\ntasks are driven by differences in the set size of the underlying memory distribution being sampled.\nThe set size of the distribution being sampled, in turn, is determined by task constraints - and ends up\nbeing symmetric when comparing single-item recognition with cued recall.\n\n5 Predicting more recognition-recall dissociations\n\nThe fact that recognition is usually easier than recall - more accurate and quicker for the same stimuli\nsets - is simply the most prominent difference between the two paradigms. Experimentalists have\nuncovered a number of interesting manipulations in memory experiments that affect performance\non these tasks differentially. These are called recognition-recall dissociations, and are prominent\nchallenges to single-process accounts of the two tasks. Why should a manipulation affect only one\ntask and not the other if they are both outcomes of the same underlying process? [20] Previous\nsingle-process accounts have had success in explaining some such dissociations. We focus here\non some that have proved relatively hard to explain without making inelegant dissociation-speci\ufb01c\nassumptions in earlier accounts [13].\n\n5.1 List strength effects and part set cuing\n\nUnidimensional strength-based models of memory like SAM and REM fail to predict the list strength\neffect [12] where participants\u2019 memory performance in free recall is lower than a controlled baseline\nfor weaker items on mixed lists (lists containing both strongly and weakly encoded items). Such\nbehavior is predicted easily by strength-based models. What they \ufb01nd dif\ufb01cult to explain is that\nperformance does not deviate from baseline in recognition tasks. The classical explanation for this\ndiscrepancy is the use of a differentiation assumption. It is assumed that stronger items are associated\nmore strongly to the encoding context, however differences between the item itself as shown, and\nits encoded image are also stronger. In free recall, this second interaction does not have an effect,\nsince the item itself is not presented, so a positive list strength effect is seen. In recognition, it is\nconjectured that the two in\ufb02uences cancel each other out, resulting in a null list strength effect [13].\nA lot of intricate assumptions have to hold for the differentiation account to hold. Our model has\na much simpler explanation for the null list-strength effect in recognition. Recognition involves\nsampling based on the strength of the associative activation of the list given a speci\ufb01c item and so\nis independent of the encoding strength of other items. On the other hand, recall involves sampling\nfrom p(item|list) across all items, in which case, having a distribution favoring other items will\nreduce the probability that the unstrengthened items will be sampled. Thus, the difference in which\nvariable the retrieval operation conditions on explains the respective presence and absence of a list\nstrength effect in recall and recognition.\nThe left panel in Figure 3 presents simulation results from our model reproducing this effect, where we\nimplement mixed lists by presenting certain stimuli more frequently during encoding and retrieve in\nthe usual manner. Hit rates are calculated for less frequently presented stimuli. For either elicitation\nmodality, the actual outcome of the retrieval itself is sampled from the appropriate conditional\n\n6\n\n\fdistribution as a speci\ufb01c item/cue. In this particular experiment, which manipulates how much\ntraining observers have on some of the items on the list, the histories entering the simulation are\ngenerated such that some items co-occur with the future retrieval cue more frequently than others, i.e.\ntwo items occur with a probability of 0.4 and 0.3 respectively, and three items occur with a probability\nof 0.1 each alongside the cue. The simulation shows a positive list strength effect for recall (weaker\nhit rates for less studied items) and a null list strength effect for recognition, congruent with data.\nOur model also reconciles the results of [1] who demonstrated that the list strength effect does not\noccur if we examine only items that are the \ufb01rst in their category to be retrieved. For category-\ninsensitive strength-based accounts, this is a serious problem. For our account, which is explicitly\nconcerned with how observers co-encode stimuli and retrieval cues, this result is no great mystery.\nFor multi-category memory tests, the presence of each semantic category instantiates a novel list\nduring encoding, such that the strength-dependent updates during retrieval apply to each individual\np(item|list) and do not apply across the other category lists.\nMore generally, the dynamic nature of the sampled distribution in our Bayesian theory accommodates\nthe theoretical views of both champions of strength-dependent activation and retrieval-dependent\nsuppression [1]. Strength-dependent activation is present in our model in the form of the Bayesian\nposterior over cue-relevant targets at the time when cued recall commences; retrieval-dependent\nsuppression of competitors is present in the form of normalization of the distribution during further\nsequential Bayesian updates as the retrieval process continues. Assigning credit differentially to\nindividual categories predicts an attenuation (though not removal) of the list strength effect, due\nto the absence of learning-induced changes for the \ufb01rst-tested items, as well diminishing memory\nperformance with testing position seen in [1].\n\nFigure 3: Reproducing (left) list strength effects and (right) the word frequency mirror effect using\nour model.\n\nThe part set cueing effect is the observation that showing participants a subset of the items to be\nrecalled during retrieval reduces their recall performance for non-shown items [11]. This effect does\nnot appear in recognition experiments, which is again problematic for unidimensional strength-based\nmemory models. Our model has a simple explanation. The presented items during retrieval are simply\ntreated as further encoding opportunities for the seen items, resulting in a list strength imbalance as\nabove. This affects recall, but not recognition for the same reasons the list strength effect does.\n\n5.2 Mirror effect\n\nAnother interesting effect that strength-based memory models have found hard to explain is the\nword-frequency mirror effect [5]. This is seen when participants see two different classes of items\nin recognition experiments. It is found, for instance, that unique items are both recognized more\naccurately as previously seen and unseen in such experiments than common items. Such a pattern of\nmemory performance is contrary to the predictions of nearly all accounts of memory that depend\non unidimensional measures of memory strength, who can only model adaptive changes in memory\nperformance via shifts in the response criterion [19] that do not permit both the hit rate and the false\nalarm rate to improve simultaneously.\n\n7\n\n00.20.40.60.81Prepend ratio00.20.40.60.81FractionRecognition HRRecognition FARRecall HRRecall FARRecognitionRecallWeak item hit rateBaselineMixed0.20.40.60.81.00\fThe essential insight of the mirror effect is that some types of stimuli are intrinsically more memorable\nthan others, a common-sense observation that has proved surprisingly dif\ufb01cult for strength-based\nmemory models to assimilate.This dif\ufb01culty extends to our own model also, but our inductive frame-\nwork allows us to express the assumptions about information that the stimuli base frequency adds\nto the picture in a clean way. Speci\ufb01cally, in our model observers use p(list|item) for recogni-\ntion, which is high for unique items and low for common items by Bayesian inversion because\np(item|list)/p(item) \u2248 1 for unique items, because they are unlikely to have been encountered\noutside the experimental context, and (cid:28) 1 for common items. In contrast, observers sample from\np(item|list) during recall, removing the effect of the frequency base rate p(item), so that the pattern\nof results is inverted: performance is equivalent or better than baseline for common stimuli than for\nrare ones [6], since they are more likely to be retrieved in general.\nThe right panel in Figure 3 shows simulation results using our model wherein we used two possible\ncues during encoding, one to test performance during retrieval and one to modify the non-retrieval\nfrequency of stimuli encounters. For this experiment, which manipulates where we have to in\ufb02uence\nhow often the relative frequency with which the observers have seen the items in task-irrelevant\ncontexts other than the retrieval task, we prepended the base case history (of size 50 time steps)\nwith differently sized prior history samples (between 10 and 50 time steps long, in steps of 5),\nwherein items co-occurred with cues that were not used during retrieval. The simulation results show\nthat, in recognition, hit rates drop and false alarm rates rise with more exposure to items outside\nthe experimental list context (high frequency items). Since our model assumes unambiguous cue\nconditioning, it predicts unchanged performance from baseline for recall. More intricate models that\npermit cue-cue associations may reproduce the advantage for common items documented empirically.\n\n6 Discussion\n\nWe have made a very simple proposal in this paper. We join multiple previous authors in arguing that\nmemory retrieval in cued recall tasks can be interpreted as a question about the likelihood of retrieving\nan item given the retrieval cue, typically the list of items given at the time of encoding [17, 8, 4].\nWe depart from previous authors in arguing that memory retrieval in item recognition tasks asks the\nprecisely opposite question: what is the likelihood of a given item having been associated with the\nlist? We integrated this insight into a simple inference-based model of memory encoding, which\nshares its formal motivations with recent inference-based models of conditioning [3, 14], and an\napproximately Bayesian model of memory retrieval, which samples memory frugally along lines\nmotivated on information-theoretic [18] and ecological grounds [16] by recent work.\nOur model is meant to be expository and ignores several large issues that other richer models typically\nengage with. For instance, it is silent about the time decay of memory particles, the partitioning of\nthe world into items and cues, and how it would go about explaining other more intricate memory\ntasks like plurality discrimination and remember-know judgments. These omissions are deliberate, in\nthe sense that we wanted to present a minimal model to deliver the core intuition behind our approach\n- that differences in memory performance in recognition and recall are attributable to no deeper\nissue than an ecological preference to test memory using more items than lists. This observation\ncan now subsequently guide and constrain the construction of more realistic models of declarative\nmemory [3]. To the extent that differences traditionally used to posit dual-process accounts of memory\ncan be accounted for using simpler models like ours, the need to proliferate neuroanatomical and\nprocess-level distinctions for various memory operations can be concomitantly reduced.\nThe distinction between recall and recognition memory also has important implications for the\npresumed architecture of machine learning systems. Modern ML systems increasingly rely on a\ncombination of distributed representation of sensory information (using deep nets) and state-centric\nrepresentation of utility information (using reinforcement learning) to achieve human-like learning\nand transfer capabilities, for example in simple Atari games [10]. The elicitation of class or category\nmembership in neural networks is quintessentially a recognition task, while the elicitation of state\nvalue functions, as well as other intermediate computations in RL are clearly recall tasks. Partly\nin realization of the large differences in the sort of memory required to support these two classes\nof learning models, researchers have taken to postulating dual-process arti\ufb01cial memories [7]. Our\ndemonstration of the fundamental unitarity of the two modes of memory performance can and should\nconstrain the design of deep RL models in simpler ways.\n\n8\n\n\fReferences\n[1] Karl-heinz B\u00e4uml. The list-strength effect: Strength-dependent competition or suppression? Psychonomic\n\nBulletin & Review, 4(2):260\u2013264, 1997.\n\n[2] Fergus IM Craik, Richard Govoni, Moshe Naveh-Benjamin, and Nicole D Anderson. The effects of divided\nattention on encoding and retrieval processes in human memory. Journal of Experimental Psychology:\nGeneral, 125(2):159, 1996.\n\n[3] Samuel J Gershman, David M Blei, and Yael Niv. Context, learning, and extinction. Psychological review,\n\n117(1):197, 2010.\n\n[4] Gary Gillund and Richard M Shiffrin. A retrieval model for both recognition and recall. Psychological\n\nreview, 91(1):1, 1984.\n\n[5] Murray Glanzer and John K Adams. The mirror effect in recognition memory: data and theory. Journal of\n\nExperimental Psychology: Learning, Memory, and Cognition, 16(1):5, 1990.\n\n[6] Vernon Gregg. Word frequency, recognition and recall. John Wiley & Sons, 1976.\n\n[7] Dharshan Kumaran, Demis Hassabis, and James L McClelland. What learning systems do intelligent agents\nneed? complementary learning systems theory updated. Trends in Cognitive Sciences, 20(7):512\u2013534,\n2016.\n\n[8] George Mandler. Recognizing: The judgment of previous occurrence. Psychological review, 87(3):252,\n\n1980.\n\n[9] James L McClelland, Bruce L McNaughton, and Randall C O\u2019reilly. Why there are complementary learning\nsystems in the hippocampus and neocortex: insights from the successes and failures of connectionist\nmodels of learning and memory. Psychological review, 102(3):419, 1995.\n\n[10] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare,\nAlex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through\ndeep reinforcement learning. Nature, 518(7540):529\u2013533, 2015.\n\n[11] Raymond S Nickerson. Retrieval inhibition from part-set cuing: A persisting enigma in memory research.\n\nMemory & cognition, 12(6):531\u2013552, 1984.\n\n[12] Roger Ratcliff, Steven E Clark, and Richard M Shiffrin. List-strength effect: I. data and discussion. Journal\n\nof Experimental Psychology: Learning, Memory, and Cognition, 16(2):163, 1990.\n\n[13] Richard M Shiffrin and Mark Steyvers. A model for recognition memory: Rem\u2014retrieving effectively\n\nfrom memory. Psychonomic bulletin & review, 4(2):145\u2013166, 1997.\n\n[14] Nisheeth Srivastava and Paul R Schrater. Classical conditioning via inference over observable situation\n\ncontexts. In Proceedings of the Annual Meeting of the Cognitive Science Society, 2014.\n\n[15] Saul Sternberg. Memory-scanning: Mental processes revealed by reaction-time experiments. American\n\nscientist, 57(4):421\u2013457, 1969.\n\n[16] Peter M Todd and Gerd Gigerenzer. Environments that make us smart: Ecological rationality. Current\n\nDirections in Psychological Science, 16(3):167\u2013171, 2007.\n\n[17] Endel Tulving and Donald M Thomson. Retrieval processes in recognition memory: Effects of associative\n\ncontext. Journal of Experimental Psychology, 87(1):116, 1971.\n\n[18] Edward Vul, Noah Goodman, Thomas L Grif\ufb01ths, and Joshua B Tenenbaum. One and done? optimal\n\ndecisions from very few samples. Cognitive science, 38(4):599\u2013637, 2014.\n\n[19] John T Wixted. Dual-process theory and signal-detection theory of recognition memory. Psychological\n\nreview, 114(1):152, 2007.\n\n[20] Andrew P Yonelinas. The nature of recollection and familiarity: A review of 30 years of research. Journal\n\nof memory and language, 46(3):441\u2013517, 2002.\n\n9\n\n\f", "award": [], "sourceid": 229, "authors": [{"given_name": "Nisheeth", "family_name": "Srivastava", "institution": "IIT Kanpur"}, {"given_name": "Edward", "family_name": "Vul", "institution": "UCSD"}]}