Kenneth J. Kurtz

The divergent autoencoder (DIVA) model of category learning

A novel theoretical approach to human category learning is proposed in which categories are represented as coordinated statistical models of the properties of the members. Key elements of the account are learning to recode inputs as task-constrained principle components and evaluating category membership in terms of model fit—that is, the fidelity of the reconstruction after recoding and decoding the stimulus. The approach is implemented as a computational model called DIVA (for DIVergent Autoencoder), an artificial neural network that uses reconstructive learning to solve N-way classification tasks. DIVA shows good qualitative fits to benchmark human learning data and provides a compelling theoretical alternative to established models.

Measuring category intuitiveness in unconstrained categorization tasks

What makes a category seem natural or intuitive? In this paper, an unsupervised categorization task was employed to examine observer agreement concerning the categorization of nine different stimulus sets. The stimulus sets were designed to capture different intuitions about classification structure. The main empirical index of category intuitiveness was the frequency of the preferred classification, for different stimulus sets. With 169 participants, and a within participants design, with some stimulus sets the most frequent classification was produced over 50 times and with others not more than two or three times. The main empirical finding was that cluster tightness was more important in determining category intuitiveness, than cluster separation. The results were considered in relation to the following models of unsupervised categorization: DIVA, the rational model, the simplicity model, SUSTAIN, an Unsupervised version of the Generalized Context Model (UGCM), and a simple geometric model based on similarity. DIVA, the geometric approach, SUSTAIN, and the UGCM provided good, though not perfect, fits. Overall, the present work highlights several theoretical and practical issues regarding unsupervised categorization and reveals weaknesses in some of the corresponding formal models.

Converging on a new role for analogy in problem solving and retrieval: when two problems are better than one

People often fail to retrieve examples analogous to a current problem or situation. There is good evidence that comparing structurally matching cases facilitates subsequent analogical access. However, current approaches offer little at the time of memory search to promote retrieval of a routinely encoded analogous source. We adapted Gick and Holyoak’s (1980, 1983) classic paradigm to investigate whether comparing two unsolved problems at test promotes retrieval of a single previously studied analogue. In Experiment 1, comparison of test problems facilitated analogical problem solving. Experiment 2 showed that comparison is the critical factor since solving two test problems separately proved ineffective. In Experiment 3, comparing two problems led to greater success for participants who read a prior analogous story than those who did not, demonstrating specifically that comparison facilitates retrieval. The three studies show that analogical access is powerfully determined by problem encoding. Implications for psychological theory and real-world applications are discussed.

Brainprint: Assessing the uniqueness, collectability, and permanence of a novel method for ERP biometrics

Abstract The human brain continually generates electrical potentials representing neural communication. These potentials can be measured at the scalp, and constitute the electroencephalogram (EEG). When the EEG is time-locked to stimulation – such as the presentation of a word – and averaged over many such presentations, the Event-Related Potential (ERP) is obtained. The functional characteristics of components of the ERP are well understood, and some components represent processing that may differ uniquely from individual to individual—such as the N400 component, which represents access to the semantic network. We applied several pattern classifiers to ERPs representing the response of individuals to a stream of text designed to be idiosyncratically familiar to different individuals. Results indicate that there are robustly identifiable features of the ERP that enable labeling of ERPs as belonging to individuals with accuracy reliably above chance (in the range of 82–97%). Further, these features are stable over time, as indicated by continued accurate identification of individuals from ERPs after a lag of up to six months. Even better, the high degree of labeling accuracy achieved in all cases was achieved with the use of only 3 electrodes on the scalp—the minimal possible number that can acquire clean data.

Comparison Promotes Learning and Transfer of Relational Categories

We investigated the effect of co-presenting training items during supervised classification learning of novel relational categories. Strong evidence exists that comparison induces a structural alignment process that renders common relational structure more salient. We hypothesized that comparisons between exemplars would facilitate learning and transfer of categories that cohere around a common relational property. The effect of comparison was investigated using learning trials that elicited a separate classification response for each item in presentation pairs that could be drawn from the same or different categories. This methodology ensures consideration of both items and invites comparison through an implicit same-different judgment inherent in making the two responses. In a test phase measuring learning and transfer, the comparison group significantly outperformed a control group receiving an equivalent training session of single-item classification learning. Comparison-based learners also outperformed the control group on a test of far transfer, that is, the ability to accurately classify items from a novel domain that was relationally alike, but surface-dissimilar, to the training materials. Theoretical and applied implications of this comparison advantage are discussed.

Detecting anomalous features in complex stimuli: the role of structured comparison.

The ability to detect anomalies in perceived stimuli is critical to a broad range of cognitive tasks, yet acquiring this ability often requires lengthy practice. In this research, we asked whether findings from research on analogical comparison can be used to aid in the acquisition of perceptual expertise. Building on findings that comparison can facilitate the detection of differences, the present research addressed two questions: (1) Does having an alignable comparison standard improve performance on a difficult detection task? (2) Can such comparison experience improve subsequent detection performance on single anomalous targets? Across 3 experiments, university undergraduates were asked to find an anomalous bone in drawings of animal skeletons. Target items including an anomaly were presented either alone or with a correct standard. Furthermore, to evaluate the impact of ease of alignment, the correct standard was presented either mirror-reversed (low alignable) or regular (high alignable). Results showed increased accuracy when a comparison standard was present and further gains when the standard was more easily alignable. In Experiment 3, we used a between-subjects design to reveal that advance comparison (as opposed to single-item training) led to improved detection of anomalies in subsequent novel examples presented as isolated targets. We conclude that the availability of a standard and ease of alignment promote encoding and processing. Furthermore, comparison-based learning confers an ongoing advantage even without standards for comparison. Therefore, task performance in application areas requiring detection of nonobvious anomalies can be improved by providing alignable standards next to targets or in advance training.

Human learning of elemental category structures: revising the classic result of Shepard, Hovland, and Jenkins (1961).

The findings of Shepard, Hovland, and Jenkins (1961) on the relative ease of learning 6 elemental types of 2-way classifications have been deeply influential 2 times over: 1st, as a rebuke to pure stimulus generalization accounts, and again as the leading benchmark for evaluating formal models of human category learning. The litmus test for models is the ability to simulate an observed advantage in learning a category structure based on an exclusive-or (XOR) rule over 2 relevant dimensions (Type II) relative to category structures that have no perfectly predictive cue or cue combination (including the linearly-separable Type IV). However, a review of the literature reveals that a Type II advantage over Type IV is found only under highly specific experimental conditions. We investigate when and why a Type II advantage exists to determine the appropriate benchmark for models and the psychological theories they represent. A series of 8 experiments link particular conditions of learning to outcomes ranging from a traditional Type II advantage to compelling non-differences and reversals (i.e., Type IV advantage). Common interpretations of the Type II advantage as either a broad-based phenomenon of human learning or as strong evidence for an attention-mediated similarity-based account are called into question by our findings. Finally, a role for verbalization in the category learning process is supported.

Re-representation in comparison: building an empirical case

Leading accounts of analogy based on structure mapping theory (Gentner 1983, 1989) give an important explanatory role to re-representation. Structural alignment is insufficiently flexible to account for human analogical processing if semantically compatible, but non-identically coded, representational elements are not permitted to match. A process of re-representation can selectively allow non-identical representational elements to be considered matches and placed in structural correspondence during comparison. However, re-representation is only a posited theoretical construct with minimal supporting evidence. An experimental paradigm called inference probing is introduced which offers a new level of empirical support for the psychological reality of re-representation. Behavioural results are presented that bear on accounts of analogy, similarity, knowledge representation and reasoning.

Observation versus classification in supervised category learning

The traditional supervised classification paradigm encourages learners to acquire only the knowledge needed to predict category membership (a discriminative approach). An alternative that aligns with important aspects of real-world concept formation is learning with a broader focus to acquire knowledge of the internal structure of each category (a generative approach). Our work addresses the impact of a particular component of the traditional classification task: the guess-and-correct cycle. We compare classification learning to a supervised observational learning task in which learners are shown labeled examples but make no classification response. The goals of this work sit at two levels: (1) testing for differences in the nature of the category representations that arise from two basic learning modes; and (2) evaluating the generative/discriminative continuum as a theoretical tool for understand learning modes and their outcomes. Specifically, we view the guess-and-correct cycle as consistent with a more discriminative approach and therefore expected it to lead to narrower category knowledge. Across two experiments, the observational mode led to greater sensitivity to distributional properties of features and correlations between features. We conclude that a relatively subtle procedural difference in supervised category learning substantially impacts what learners come to know about the categories. The results demonstrate the value of the generative/discriminative continuum as a tool for advancing the psychology of category learning and also provide a valuable constraint for formal models and associated theories.

Category learning in the context of co-presented items

A series of four studies explore how the presentation of multiple items on each trial of a categorization task affects the course of category learning. In a three-category supervised classification task involving multi-dimensionally varying artificial organism-like stimuli, learners are shown a target plus two context items on every trial, with the context items’ category membership explicitly identified. These triads vary in whether one, two, or all three categories are represented. This presentation context can support within-category comparison and/or between-category contrast. The most successful learning occurs when all categories are represented in each trial. This pattern occurs across two different underlying category structures and across variations in learners’ prior knowledge of the relationship between the target and context items. These results appear to contrast with some other recent findings and make clear the potential importance of context-based inter-item evaluation in human category learning, which has implications for psychological theory and for real-world learning environments.