Search

We introduce the Category Abstraction Learning (CAL) model, a cognitive framework formally describing category learning built on similarity-based generalization, dissimilarity-based abstraction, two attention learning mechanisms, error-driven knowledge structuring, and stimulus memorization. Our hypotheses draw on an array of empirical and theoretical insights connecting reinforcement and category learning. The key novelty of the model is its explanation of how rules are learned from scratch based on three central assumptions. (a) Category rules emerge from two processes of stimulus generalization (similarity) and its direct inverse (category contrast) on independent dimensions. (b) Two attention mechanisms guide learning by focusing on rules, or on the contexts in which they produce errors. (c) Knowing about these contexts inhibits executing the rule, without correcting it, and consequently leads to applying partial rules in different situations. The model is designed to capture both systematic and individual differences in a broad range of learning paradigms. We illustrate the model's explanatory scope by simulating several benchmarks, including the classic Six Problems, the 5-4 problem, and linear separability. Beyond the common approach of predicting average response probabilities, we also propose explanations for more recently studied phenomena that challenge existing learning accounts, regarding task instructions, individual differences in rule extrapolation in three different tasks, individual attention shifts to stimulus features during learning, and other phenomena. We discuss CAL's relation to different models, and its potential to measure the cognitive processes regarding attention, abstraction, error detection, and memorization from multiple psychological perspectives. (PsycInfo Database Record (c) 2021 APA, all rights reserved).

Relative to studying alone, guessing the meanings of unknown words can improve later recognition of their meanings, even if those guesses were incorrect – the pretesting effect (PTE). The error-correction hypothesis suggests that incorrect guesses produce error signals that promote memory for the meanings when they are revealed. The current research sought to test the error-correction explanation of the PTE. In three experiments, participants studied unfamiliar Finnish-English word pairs by either studying each complete pair or by guessing the English translation before its presentation. In the latter case, the participants also guessed which of two categories the word belonged to. Hence, guesses from the correct category were semantically closer to the true translation than guesses from the incorrect category. In Experiment 1, guessing increased subsequent recognition of the English translations, especially for translations that were presented on trials in which the participants’ guesses were from the correct category. Experiment 2 replicated these target recognition effects while also demonstrating that they do not extend to associative recognition performance. Experiment 3 again replicated the target recognition pattern, while also examining participants’ metacognitive recognition judgments. Participants correctly judged that their memory would be better after small than after large errors, but incorrectly believed that making any errors would be detrimental, relative to study-only. Overall, the data are inconsistent with the error-correction hypothesis; small, within-category errors produced better recognition than large, cross-category errors. Alternative theories, based on elaborative encoding and motivated learning, are considered.

It is generally assumed that the Rescorla and Wagner (1972) model adequately accommodates the full results of simple cue competition experiments in humans (e.g. Dickinson et al., 1984), while the Bush and Mosteller (1951) model cannot. We present simulations that demonstrate this assumption is wrong in at least some circumstances. The Rescorla-Wagner model, as usually applied, fits the full results of a simple forward cue-competition experiment no better than the Bush-Mosteller model. Additionally, we present a novel finding, where letting the associative strength of all cues start at an intermediate value (rather than zero), allows this modified model to provide a better account of the experimental data than the (equivalently modified) Bush-Mosteller model. This modification also allows the Rescorla-Wagner model to account for a redundancy effect experiment (Uengoer et al., 2013); something that the unmodified model is not able to do. Furthermore, the modified Rescorla-Wagner model can accommodate the effect of varying the proportion of trials on which the outcome occurs (i.e. the base rate) on the redundancy effect (Jones et al., 2019). Interestingly, the initial associative strength of cues varies in line with the outcome base rate. We propose that this modification provides a simple way of mathematically representing uncertainty about the causal status of novel cues within the confines of the Rescorla-Wagner model. The theoretical implications of this modification are discussed. We also briefly introduce free and open resources to support formal modelling in associative learning. Keywords: associative learning, prediction error, uncertainty, modelling, blocking, redundancy effect, open science.

Attempting to retrieve the answer to a question on an initial test can improve memory for that answer on a subsequent test, relative to an equivalent study period. Such retrieval attempts can be beneficial even when they are unsuccessful, although this benefit is usually only seen with related word pairs. Three experiments examined the effects of pretesting for both related (e.g., pond-frog) and unrelated (e.g., pillow-leaf) word pairs on cued recall and target recognition. Pretesting improved subsequent cued recall performance for related but not for unrelated word pairs, relative to simply studying the word pairs. Tests of target recognition, by contrast, revealed benefits of pretesting for memory of targets from both related and unrelated word pairs. These data challenge popular theories that suggest that the pretesting effect depends on partial activation of the target during the pretesting phase.

Analogical transfer has been previously reported to occur between rule-based, but not information-integration, perceptual category structures (Casale, Roeder, & Ashby, 2012). The current study investigated whether a similar pattern of results would be observed in cross-modality transfer. Participants were trained on either a rule-based structure, or an information-integration structure, using visual stimuli. They were then tested on auditory stimuli that had the same underlying abstract category structure. Transfer performance was assessed relative to a control group who did not receive training on the visual stimuli. No cross-modality transfer was found, irrespective of the category structure employed.

Three experiments were conducted to investigate a possible role for certainty in human causal learning. In these experiments, human participants were initially trained with a set of cues, each of which was followed by the presence or absence of an outcome. In a subsequent training stage, 2 of these cues were trained in a causal compound, and the change in associative strength for each of the cues was compared, using a procedure based on Rescorla (2001). In each experiment, the cues differed in both their causal certainty (on the part of participants) and size of their prediction error (with respect to the outcome). The cue with the larger prediction error was always the cue with the more certain causal status. According to established prediction error models (Bush & Mosteller, 1951; Rescorla, 2001; Rescorla & Wagner, 1972), a larger prediction error should result in a greater updating of associative strength. However, the opposite was observed, as participants always learned more about the cue with the smaller prediction error. A plausible explanation is that participants engaged in a form of theory protection, in which they were resistant to updating their existing beliefs about cues with a certain causal status. Instead, participants appeared to attribute outcomes to cues with a comparatively uncertain causal status, in an apparent violation of prediction error. The potential role of attentional processes (Mackintosh, 1975; Pearce & Hall, 1980) in explaining these results is also discussed. (PsycInfo Database Record (c) 2020 APA, all rights reserved).

Theories of associative learning often propose that learning is proportional to prediction error, or the difference between expected events and those that occur. Spicer et al. (2020) suggested an alternative, that humans might instead selectively attribute surprising outcomes to cues that they are not confident about, to maintain cue-outcome associations about which they are more confident. Spicer et al. reported three predictive learning experiments, the results of which were consistent with their proposal ("theory protection") rather than a prediction error account (Rescorla, 2001). The four experiments reported here further test theory protection against a prediction error account. Experiments 3 and 4 also test the proposals of Holmes et al. (2019), who suggested a function mapping learning to performance that can explain Spicer et al.'s results using a prediction-error framework. In contrast to the previous study, these experiments were based on inhibition rather than excitation. Participants were trained with a set of cues (represented by letters), each of which was followed by the presence or absence of an outcome (represented by + or -). Following this, a cue that previously caused the outcome (A+) was placed in compound with another cue (B) with an ambiguous causal status (e.g., a novel cue in Experiment 1). This compound (AB-) did not cause the outcome. Participants always learned more about B in the second training phase, despite A always having the greater prediction error. In Experiments 3 and 4, a cue with no apparent prediction error was learned about more than a cue with a large prediction error. Experiment 4 tested participants' relative confidence about the causal status of cues A and B prior to the AB- stage, producing findings that are consistent with theory protection and inconsistent with the predictions of Rescorla, and Holmes et al. (PsycInfo Database Record (c) 2022 APA, all rights reserved).

The Inverse Base Rate effect (IBRE; Medin & Edelson, 1988) is a non-rational behavioural phenomenon in predictive learning. Canonically, participants learn that the AB stimulus compound leads to one outcome and that AC leads to another outcome, with AB being presented three times as often as AC. When subsequently presented with BC, the outcome associated with AC is preferentially selected, in opposition to the underlying base rates of the outcomes. An error-driven learning account (Kruschke, 2001b) is the leading current explanation of the IBRE. A key component of this account is prediction error, a concept previously linked to a number of brain areas including the anterior cingulate, the striatum and the dorsolateral prefrontal cortex. The present work is the first fMRI study to directly examine the IBRE. Activations were noted in brain areas linked to prediction error, including the caudate body, the anterior cingulate, the ventromedial prefrontal cortex and the right dorsolateral prefrontal cortex. Analysing the difference in activations for singular key stimuli (B and C), as well as frequency matched controls, supports the predictions made by the error-driven learning account.

Polymorphous concepts are hard to learn, and this is perhaps surprising because they, like many natural concepts, have an overall similarity structure. However, the dimensional summation hypothesis (Milton and Wills Journal of Experimental Psychology: Learning, Memory and Cognition, 30, 407–415 2004) predicts this difficulty. It also makes a number of other predictions about polymorphous concept formation, which are tested here. In Experiment 4, we confirm the theory’s prediction that polymorphous concept formation should be facilitated by deterministic pretraining on the constituent features of the stimulus. This facilitation is relative to an equivalent amount of training on the polymorphous concept itself. In further experiments, we compare the predictions of the dimensional summation hypothesis with a more general strategic account (Experiment 2), a seriality of training account (Experiment 3), a stimulus decomposition account (also Experiment 3), and an error-based account (Experiment 4). The dimensional summation hypothesis provides the best account of these data. In Experiment 5, a further prediction is confirmed—the single feature pretraining effect is eliminated by a concurrent counting task. The current experiments suggest the hypothesis that natural concepts might be acquired by the deliberate serial summation of evidence. This idea has testable implications for classroom learning.

The current research examined the effects of errorful generation on memory, focusing particularly on the roles of motivation and surprise. In two experiments, participants were first presented with photographs of faces and were asked to associate four facts with each photograph. On Generate trials, the participants guessed two of the facts (Guess targets) before those correct facts, and another two correct facts (Study targets), were revealed. On the remaining Read trials, all four facts were presented without a guessing stage. In Experiment 1, participants also ranked their motivation to know the answers before they were revealed, or their surprise on learning the true answers. Guess targets were subsequently better recognised than the concurrently presented, non-guessed Study targets. Guess targets were also better recognised than Read targets, and recognition of Study and Read targets did not differ. Errorful generation also increased self-reported motivation, but not surprise. Experiment 2 showed that the results of Experiment 1 can outlive a 20-minute delay, and that they generalise to a more challenging recognition test. Together, the results suggest that errorful generation improves memory specifically for the guessed fact, and this may be linked to an increase in motivation to learn that fact.

Spicer et al. (2020) reported a series of causal learning experiments in which participants appeared to learn most readily about cues when they were not certain of their causal status and proposed that their results were a consequence of participants' use of theory protection. In the present issue, Chan et al. (2021) present an alternative view, using a modification of Rescorla and Wagner's (1972) influential model of learning. Although the explanation offered by Chan et al. appears very different from that suggested by Spicer et al., there are conceptual commonalities. Here we briefly discuss the similarities and differences of the 2 approaches and agree with Chan et al.'s proposal that the best way to advance the debate will be to test situations in which the 2 theories make differing predictions. (PsycInfo Database Record (c) 2021 APA, all rights reserved).

Integral stimuli (e.g. colours varying in saturation and brightness) are classically considered to be processed holistically (i.e. as undifferentiated stimulus wholes); people analyze such stimuli into their consistent dimensions only with substantial time, effort, training, or instruction (Foard & Kemler Nelson, 1984). In contrast, Combination Theory (Wills et al., 2015) argues that the dimensions of integral stimuli are quickly combined. Through an investigation of the effects of time pressure, we support Combination Theory over the classical holistic-to-analytic account. Specifically, using colored squares varying in saturation and brightness, we demonstrate that the prevalence of single-dimension classification increases as stimulus presentation time is reduced. We conclude that integral stimuli are not slowly analyzed, they are quickly synthesized.

Although instructions often emphasize categories (e.g., odd number→left hand response) rather than specific stimuli (e.g., 3→left hand response), learning is often interpreted in terms of stimulus-response (S-R) bindings or, less frequently, stimulus-classification (S-C) bindings with little attention being paid to the importance of category-response (C-R) bindings. In a training-transfer paradigm designed to investigate the early stages of category learning, participants were required to classify stimuli according to the category templates presented prior to each block (Experiments 1-4). In some transfer blocks the stimuli, categories and/or responses could be novel or repeated from the preceding training phase. Learning was assessed by comparing the transfer-training performance difference across conditions. Participants were able to rapidly transfer C-R associations to novel stimuli but evidence of S-C transfer was much weaker and S-R transfer was largely limited to conditions where the stimulus was classified under the same category. Thus, even though there was some evidence that learned S-R and S-C associations contributed to performance, learned C-R associations seemed to play a much more important role. In a final experiment (Experiment 5) the stimuli themselves were presented prior to each block, and the instructions did not mention the category structure. In this experiment, the evidence for S-R learning outweighed the evidence for C-R learning, indicating the importance of instructions in learning. The implications for these findings to the learning, cognitive control, and automaticity literatures are discussed.

In the phenomenon of transfer along a continuum (TAC), initial training on easy items facilitates later learning of a harder discrimination. TAC is a widely replicated cross-species phenomenon that is well predicted by certain kinds of associative theory. A recent report of an approximately opposite phenomenon (i.e., facilitation by initial training on hard items) poses a puzzle for such theories, but is predicted by a dual-system model (COVIS). However, across four experiments, we present substantial evidence that this counterintuitive finding was in error. Rather, the result appears to be a false positive and, as such, should not form part of the evidence base for COVIS nor be considered as a counter-example to the pervasive TAC phenomenon.

The Inverse Base Rate Effect (IBRE; Medin and Edelson (1988)) is a non-rational behavioural phenomenon in predictive learning. In the IBRE, participants learn that a stimulus compound AB leads to one outcome and that another compound AC leads to a different outcome. Importantly, AB and its outcome are presented three times as often as AC (and its outcome). On test, when asked which outcome to expect on presentation of the novel compound BC, participants preferentially select the rarer outcome, previously associated with AC. This is irrational because, objectively, the common outcome is more likely. Usually, the IBRE is attributed to greater attention paid to cue C than to cue B, and so is an excellent test for attentional learning models. The current experiment tested a simple model of attentional learning proposed by Le Pelley, Mitchell, Beesley, George, and Wills (2016) where attention paid to a stimulus is determined by its associative strength. This model struggles to capture the IBRE, but a potential solution suggested by the authors appeals to the role of experimental context. In the present paper, we derive three predictions from their account concerning the effect of changing to a novel experimental context at test, and examine these predictions empirically. Only one of the predictions was supported, concerning the effect of a context shift on responding to a novel cue, was supported. In contrast, Kruschke (2001b)’s EXIT model, in which attention and associative strength can vary independently, captured the data with a high degree of quantitative accuracy.

Comparative and cognitive psychologists interpret performance in different ways. Animal researchers invoke a dominant construct of associative learning. Human researchers acknowledge humans’ capacity for explicit-declarative cognition. This article offers a way to bridge a divide that defeats productive cross-talk. We show that animals often challenge the associative-learning construct, and that it does not work to try to stretch the associative-learning construct to encompass these performances. This approach thins and impoverishes that important construct. We describe an alternative approach that restrains the construct of associative learning by giving it a clear operational definition. We apply this approach in several comparative domains to show that different task variants change—in concert—the level of awareness, the declarative nature of knowledge, the dimensional breadth of knowledge, and the brain systems that organize learning. These changes reveal dissociable learning processes that a unitary associative construct cannot explain but a neural-systems framework can explain. These changes define the limit of associative learning and the threshold of explicit cognition. The neural-systems framework can broaden empirical horizons in comparative psychology. It can offer animal models of explicit cognition to cognitive researchers and neuroscientists. It can offer simple behavioral paradigms for exploring explicit cognition to developmental researchers. It can enliven the synergy between human and animal research, promising a productive future for both.

Two experiments are reported that demonstrate rate of learning in an allergy prediction task can b e a ffected b y the predictive history of the c ues involved, even if that history relates to ou tcomes different t o those being currently learned about. Predictive history is defined here as a cue’s prior status as either a good or a poor predictor of outcomes. Our r esults are problematic for commonly employed associative theories of human contingency learning but also p rovide e vidence for the sort of associability-change process envisaged by the Mackintosh (1975) theory.

publisher: Elsevier articletitle: In Defence of Effect-Centric Research journaltitle: Journal of Applied Research in Memory and Cognition articlelink: http://dx.doi.org/10.1016/j.jarmac.2016.10.005 content_type: simple-article copyright: © 2016 Society for Applied Research in Memory and Cognition. Published by Elsevier Inc. All rights reserved.

OBJECTIVE Generalization is the application of existing knowledge to novel situations. Questions remain about the precise role of the hippocampus in this facet of learning, but a connectionist model by Gluck and Myers (1993) predicts that generalization should be enhanced following hippocampal damage. METHOD In a two-category learning task, a group of amnesic patients (n = 9) learned the training items to a similar level of accuracy as matched controls (n = 9). Both groups then classified new items at various levels of distortion. RESULTS The amnesic group showed significantly more accurate generalization to high-distortion novel items, a difference also present compared to a larger group of unmatched controls (n = 33). CONCLUSIONS The model prediction of a broadening of generalization gradients in amnesia, at least for items near category boundaries, was supported by the results. Our study shows for the first time that amnesia can sometimes improve generalization. (PsycINFO Database Record

This article presents a comprehensive survey of research concerning interactions between associative learning and attention in humans. Four main findings are described. First, attention is biased toward stimuli that predict their consequences reliably (learned predictiveness). This finding is consistent with the approach taken by Mackintosh (1975) in his attentional model of associative learning in nonhuman animals. Second, the strength of this attentional bias is modulated by the value of the outcome (learned value). That is, predictors of high-value outcomes receive especially high levels of attention. Third, the related but opposing idea that uncertainty may result in increased attention to stimuli (Pearce & Hall, 1980), receives less support. This suggests that hybrid models of associative learning, incorporating the mechanisms of both the Mackintosh and Pearce-Hall theories, may not be required to explain data from human participants. Rather, a simpler model, in which attention to stimuli is determined by how strongly they are associated with significant outcomes, goes a long way to account for the data on human attentional learning. The last main finding, and an exciting area for future research and theorizing, is that learned predictiveness and learned value modulate both deliberate attentional focus, and more automatic attentional capture. The automatic influence of learning on attention does not appear to fit the traditional view of attention as being either goal-directed or stimulus-driven. Rather, it suggests a new kind of “derived” attention.

The influential competition between verbal and implicit systems (COVIS) model proposes that category learning is driven by two competing neural systems-an explicit, verbal, system, and a procedural-based, implicit, system. In the current fMRI study, participants learned either a conjunctive, rule-based (RB), category structure that is believed to engage the explicit system, or an information-integration category structure that is thought to preferentially recruit the implicit system. The RB and information-integration category structures were matched for participant error rate, the number of relevant stimulus dimensions, and category separation. Under these conditions, considerable overlap in brain activation, including the prefrontal cortex, basal ganglia, and the hippocampus, was found between the RB and information-integration category structures. Contrary to the predictions of COVIS, the medial temporal lobes and in particular the hippocampus, key regions for explicit memory, were found to be more active in the information-integration condition than in the RB condition. No regions were more activated in RB than information-integration category learning. The implications of these results for theories of category learning are discussed. Hum Brain Mapp 37:3557-3574, 2016. © 2016 Wiley Periodicals, Inc.

Inhibitory processes have been implicated in depressive rumination. Inhibitory deficits may cause difficulties in disengaging from ruminative content (e.g., Joormann, 2005), or rumination may constitute a working memory load, causing deficits in inhibitory control (e.g., Hertel, 2004). These hypotheses have different implications for the treatment of depression. We conducted a systematic review of existing evidence, and conclude that most studies do not unambiguously measure inhibition. The majority of published evidence is correlational, and thus supports neither causal direction. No published experimental studies have investigated the inhibitory deficit -? rumination causal direction, and only six have investigated the rumination -? inhibitory deficit hypothesis. In two of these studies the dependent variable has low construct validity. One study reported no effect of rumination on interference, and three did not control for mood effects. There is need for carefully designed experimental research that has the potential to investigate these proposed causal mechanisms.

Potts and Shanks (2014) recently reported that making mistakes improved the encoding of novel information compared with simply studying. This benefit of generating errors is counterintuitive, since it resulted in less study time and more opportunity for proactive interference. Five experiments examined the effect of generating errors versus studying on item recognition, cued recall, associative recognition, two-alternative forced choice and multiple-choice performance. Following Potts and Shanks (2014), participants first attempted to learn the English definitions of either very rare English words or Euskara nouns. During encoding, participants either guessed the definition (and almost always made an error) before the correct definition was revealed, or simply studied the words for an equivalent period. Experiments 1–4 used rare English words. In these experiments, generating errors led to better subsequent recognition of both the cues and targets compared with studying (Experiments 1 and 3). Tests of cued recall and associative recognition, by contrast, revealed no significant benefit of generating errors over studying (Experiments 1–3). Generating errors during encoding also improved performance on a two-alternative forced choice test when the correct target was presented with a novel foil, but not when the familiarity of the target and the foil was matched (Experiment 4). In Experiment 5, a different set of materials – Euskara nouns – and a different (intermixed) encoding procedure was adopted. Here, guessing improved target recognition (performance was improved on a multiple-choice test with unfamiliar foils), but impaired cued recall performance. These results suggest that, when learning word pairs that do not have a pre-existing semantic association, generating errors strengthens the cues and targets in isolation, but does not strengthen the cue-target associations.

The extent to which human outcome–response (O-R) priming effects are automatic or under cognitive control is currently unclear. Two experiments tested the effect of cognitive load on O-R priming to shed further light on the debate. In Experiment 1, two instrumental responses earned beer and chocolate points in an instrumental training phase. Instrumental response choice was then tested in the presence of beer, chocolate, and neutral stimuli. On test, a Reversal instruction group was told that the stimuli signalled which response would not be rewarded. The transfer test was also conducted under either minimal (No Load) or considerable (Load) cognitive load. The Non-Reversal groups showed O-R priming effects, where the reward cues increased the instrumental responses that had previously produced those outcomes, relative to the neutral stimulus. This effect was observed even under cognitive load. The Reversal No Load group demonstrated a reversed effect, where response choice was biased towards the response that was most likely to be rewarded according to the instruction. Most importantly, response choice was at chance in the Reversal Load condition. In Experiment 2, cognitive load abolished the sensitivity to outcome devaluation that was otherwise seen when multiple outcomes and responses were cued on test. Collectively, the results demonstrate that complex O-R priming effects are sensitive to cognitive load, whereas the very simple, standard O-R priming effect is more robust.

Two experiments investigated what makes it more likely that pigeons' behavior will come under the control of multiple relevant visual stimulus dimensions. Experiment 1 investigated the effect of stimulus set structure, using a conditional discrimination between circles that differed in both hue and diameter. Two training conditions differed in whether hue and diameter were correlated in the same way within positive and negative stimulus sets as between sets. Transfer tests showed that all pigeons came under the control of both dimensions, regardless of stimulus set structure. Experiment 2 investigated the effect of the relative salience of the stimulus differences on three visual dimensions. Pigeons learned a multiple simultaneous discrimination between circular patches of sinusoidal gratings that differed in hue, orientation, and spatial frequency. In initial training, each stimulus only included one positive or negative feature, and the stimulus differences on the three dimensions were adjusted so that the rates of learning about the three dimensions were kept approximately equal. Transfer tests showed that all three dimensions acquired control over behavior, with no single dimension dominating consistently across pigeons. Subsequently the pigeons were trained in a polymorphous category discrimination using all three dimensions, and the level of control by the three dimensions tended to become more equal as polymorphous training continued. We conclude that the salience of the stimulus differences on different dimensions is an important factor in whether pigeons will come under the control of multiple dimensions of visual stimuli. (PsycINFO Database Record

Does cognition begin with an undifferentiated stimulus whole, which can be divided into distinct attributes if time and cognitive resources allow (Differentiation Theory)? Or does it begin with the attributes, which are combined if time and cognitive resources allow (Combination Theory)? Across psychology, use of the terms analytic and non-analytic imply that Differentiation Theory is correct—if cognition begins with the attributes, then synthesis, rather than analysis, is the more appropriate chemical analogy. We re-examined four classic studies of the effects of time pressure, incidental training, and concurrent load on classification and category learning (Kemler Nelson, 1984; Smith & Kemler Nelson, 1984; Smith & Shapiro, 1989; Ward, 1983). These studies are typically interpreted as supporting Differentiation Theory over Combination Theory, while more recent work in classification (Milton et al., 2008, et seq.) supports the opposite conclusion. Across seven experiments, replication and re-analysis of the four classic studies revealed that they do not support Differentiation Theory over Combination Theory—two experiments support Combination Theory over Differentiation Theory, and the remainder are compatible with both accounts. We conclude that Combination Theory provides a parsimonious account of both classic and more recent work in this area. The presented data do not require Differentiation Theory, nor a Combination–Differentiation hybrid account.

The effects of two different types of training on rule-based and information-integration category learning were investigated in two experiments. In observational training, a category label is presented, followed by an example of that category and the participant's response. In feedback training, the stimulus is presented, and the participant assigns it to a category and then receives feedback about the accuracy of that decision. Ashby, Maddox, and Bohil (2002. Observational versus feedback training in rule-based and information-integration category learning. Memory & Cognition, 30, 666–677) reported that feedback training was superior to observational training when learning information-integration category structures, but that training type had little effect on the acquisition of rule-based category structures. These results were argued to support the COVIS (competition between verbal and implicit systems) dual-process account of category learning. However, a number of nonessential differences between their rule-based and information-integration conditions complicate interpretation of these findings. Experiment 1 controlled between-category structures for participant error rates, category separation, and the number of stimulus dimensions relevant to the categorization. Under these more controlled conditions, rule-based and information-integration category structures both benefited from feedback training to a similar degree. Experiment 2 maintained this difference in training type when learning a rule-based category that had otherwise been matched, in terms of category overlap and overall performance, with the rule-based categories used in Ashby et al. These results indicate that differences in dimensionality between the category structures in Ashby et al. is a more likely explanation for the interaction between training type and category structure than the dual-system explanation that they offered.