Statistical Learning of Unfamiliar Sounds as Trajectories Through a Perceptual Similarity Space

In typical statistical learning studies, researchers define sequences in terms of the probability of the next item in the sequence given the current item (or items), and they show that high probability sequences are treated as more familiar than low probability sequences. Existing accounts of these phenomena all assume that participants represent statistical regularities more or less as they are defined by the experimenters—as sequential probabilities of symbols in a string. Here we offer an alternative, or possibly supplementary, hypothesis. Specifically, rather than identifying or labeling individual stimuli discretely in order to predict the next item in a sequence, we need only assume that the participant is able to represent the stimuli as evincing particular similarity relations to one another, with sequences represented as trajectories through this similarity space. We present experiments in which this hypothesis makes sharply different predictions from hypotheses based on the assumption that sequences are learned over discrete, labeled stimuli. We also present a series of simulation models that encode stimuli as positions in a continuous two‐dimensional space, and predict the next location from the current location. Although no model captures all of the data presented here, the results of three critical experiments are more consistent with the view that participants represent trajectories through similarity space rather than sequences of discrete labels under particular conditions.     

Common and distinctive features in stimulus similarity: A modified version of the contrast model

Featural representations of similarity data assume that people represent stimuli in terms of a set of discrete properties. In this article, we consider the differences in featural representations that arise from making four different assumptions about how similarity is measured. Three of these similarity models— the common features model, the distinctive features model, and Tversky’s seminal contrast model—have been considered previously. The other model is new and modifies the contrast model by assuming that each individual feature only ever acts as a common or distinctive feature. Each of the four models is tested on previously examined similarity data, relating to kinship terms, and on a new data set, relating to faces. In fitting the models, we have used the geometric complexity criterion to balance the competing demands of data-fit and model complexity. The results show that both common and distinctive features are important for stimulus representation, and we argue that the modified contrast model combines these two components in a more effective and interpretable way than Tversky’s original formulation.     

Generalization and similarity in exemplar models of categorization: Insights from machine learning

Exemplar theories of categorization depend on similarity for explaining subjects’ ability to generalize to new stimuli. A major criticism of exemplar theories concerns their lack of abstraction mechanisms and thus, seemingly, of generalization ability. Here, we use insights from machine learning to demonstrate that exemplar models can actually generalize very well. Kernel methods in machine learning are akin to exemplar models and are very successful in real-world applications. Their generalization performance depends crucially on the chosen similarity measure. Although similarity plays an important role in describing generalization behavior, it is not the only factor that controls generalization performance. In machine learning, kernel methods are often combined with regularization techniques in order to ensure good generalization. These same techniques are easily incorporated in exemplar models. We show that the generalized context model (Nosofsky, 1986) and ALCOVE (Kruschke, 1992) are closely related to a statistical model called kernel logistic regression. We argue that generalization is central to the enterprise of understanding categorization behavior, and we suggest some ways in which insights from machine learning can offer guidance.     

A computational feature binding model of human texture perception

We present a computational model for human texture perception which assigns functional principles to the Gestalt laws of similarity and proximity. Motivated by early vision mechanisms, in the first stage, local texture features are extracted by utilizing multi-scale filtering and nonlinear spatial pooling. In the second stage, features are grouped according to the spatial feature binding model of the competitive layer model (CLM; Wersing et al. 2001). The CLM uses cooperative and competitive interactions in a recurrent network, where binding is expressed by the layer-wise coactivation of feature-representing neurons. The Gestalt law of similarity is expressed by a non-Euclidean distance measure in the abstract feature space with proximity being taken into account by a spatial component. To choose the stimulus dimensions which allow the most salient similarity-based texture segmentation, the feature similarity metrics is reduced to the directions of maximum variance. We show that our combined texture feature extraction and binding model performs segmentation in strong conformity with human perception. The examples range from classical microtextures and Brodatz textures to other classical Gestalt stimuli, which offer a new perspective on the role of texture for more abstract similarity grouping.     

Feature selection for inductive generalization

Judging similarities among objects, events, and experiences is one of the most basic cognitive abilities, allowing us to make predictions and generalizations. The main assumption in similarity judgment is that people selectively attend to salient features of stimuli and judge their similarities on the basis of the common and distinct features of the stimuli. However, it is unclear how people select features from stimuli and how they weigh features. Here, we present a computational method that helps address these questions. Our procedure combines image-processing techniques with a machine-learning algorithm and assesses feature weights that can account for both similarity and categorization judgment data. Our analysis suggests that a small number of local features are particularly important to explain our behavioral data.     

Rating the similarity of simple perceptual stimuli: asymmetries induced by manipulating exposure frequency.

When judging the similarity of two stimuli, people's ratings often differ depending on the order in which the comparison is presented (A vs. B or B vs. A). Such directional asymmetries have typically been demonstrated using complex concepts that have a large number of semantic features and a standard explanation is that different sets of features are emphasized depending on the direction of the comparison. In this study, we show that directional asymmetries in the similarity of simple perceptual stimuli can be predictably manipulated merely by presenting each member of a pair with different frequency. Participants rated the similarity of color patches before and after performing an irrelevant training task in which a subset of colors was presented ten times more frequently than others. The similarity ratings after training were significantly more asymmetric than the ratings before training. We discuss the implications of these findings for models of similarity judgment and propose a computationally explicit explanation based on asymmetries in representational stability.     

Relational similarity and the nonindependence of features in similarity judgments

Four experiments examined the hypothesis that simple attributional features and relational features operate differently in the determination of similarity judgments. Forced choice similarity judgments ("Is X or Y more similar to Z?") and similarity rating tasks demonstrate that making the same featural change in two geometric stimuli unequally affects their judged similarity to a third stimulus (the comparison stimulus). More specifically, a featural change that causes stimuli to be more superficially similar and less relationally similar increases judged similarity if it occurs in stimuli that already share many superficial attributes, and decreases similarity if it occurs in stimuli that do not share as many superficial attributes. These results argue against an assumption of feature independence which asserts that the degree to which a feature shared by two objects affects similarity is independent of the other features shared by the objects. The MAX hypothesis is introduced, in which attributional and relational similarities are separately pooled, and shared features affect similarity more if the pool they are in is already relatively large. The results support claims that relations and attributes are psychologically distinct and that formal measures of similarity should not treat all types of matching features equally.     

Measures of similarity in models of categorization

This paper concerns the use of similarities based on geometric distance in models of categorization. Two problematic implications of such similarities are outlined. First, in a comparison between two stimuli, geometric distance implies that matching features are not taken into account. Second, missing features are assumed not to exist. Only nonmatching features enter into calculations of similarity. A new model is constructed that is based on the ALCOVE model (Kruschke, 1992), but it uses a feature-matching similarity measure (see, e.g., Tversky, 1977) rather than a geometric one. It is an on-line model in the sense that both dimensions and exemplars are constructed during the categorization process. The model accounts better than ALCOVE does for data with missing features (Experiments 1 and 2) and at least as well as ALCOVE for a data set without missing features (Nosofsky, Kruschke, & McKinley, 1992). This suggests that, at least for some stimulus materials, similarity in categorization is more akin to a feature-matching procedure than to geometric distance calculation.     

Memory-conjunction errors: Miscombination of stored stimulus features can produce illusions of memory

We demonstrate that subjects will often claim to have previously seen a new stimulus if they have previously seen stimuli containing its component features. Memory for studied stimuli was measured using a "yes"/"no" recognition test. There were three types of test stimuli: target stimuli, which had been presented during study, conjunction stimuli, constructed by combining the features of separate study stimuli, and feature stimuli, in which studied stimulus features were combined with new, unstudied, features. For both nonsense words and faces, the subjects made many more false alarms for conjunction than for feature stimuli. Additional experiments demonstrated that the results were not due to physical similarity between study and test stimuli and that conjunction errors were much more common than feature errors in recall. The results demonstrate that features of stored stimuli maintain some independence in memory and can be incorrectly combined to produce recognition errors.     

Neural theory for the perception of causal actions

The efficient prediction of the behavior of others requires the recognition of their actions and an understanding of their action goals. In humans, this process is fast and extremely robust, as demonstrated by classical experiments showing that human observers reliably judge causal relationships and attribute interactive social behavior to strongly simplified stimuli consisting of simple moving geometrical shapes. While psychophysical experiments have identified critical visual features that determine the perception of causality and agency from such stimuli, the underlying detailed neural mechanisms remain largely unclear, and it is an open question why humans developed this advanced visual capability at all. We created pairs of naturalistic and abstract stimuli of hand actions that were exactly matched in terms of their motion parameters. We show that varying critical stimulus parameters for both stimulus types leads to very similar modulations of the perception of causality. However, the additional form information about the hand shape and its relationship with the object supports more fine-grained distinctions for the naturalistic stimuli. Moreover, we show that a physiologically plausible model for the recognition of goal-directed hand actions reproduces the observed dependencies of causality perception on critical stimulus parameters. These results support the hypothesis that selectivity for abstract action stimuli might emerge from the same neural mechanisms that underlie the visual processing of natural goal-directed action stimuli. Furthermore, the model proposes specific detailed neural circuits underlying this visual function, which can be evaluated in future experiments.     

Visual judgments of kinship: an alternative perspective

Following other researchers, we investigated the premise that visual judgment of kinship might be modelled as a signal-detection task, strictly related to similar facial features. We measured subjects' response times to face-pair stimuli while they performed visual judgments of kinship, similarity, or dissimilarity, and examined some priming effects involved. Our results show that kinship judgment takes longer on average than either similarity or dissimilarity judgment-which is compatible with existing models, yet might also suggest that kinship judgments are of a more complex character. In our priming study we observed selective suppression/enhancement of the efficacy of dissimilarity judgments whenever they followed similarity and kinship judgments. This finding confounds the notion, inherent in previous models, of resemblance cues signalling for kinship, since similarity and dissimilarity cannot be considered just as opposite concepts, and observed priming effects need to be explicitly modelled, including dissimilarity cues. To model kinship judgments across faces that are perceived as dissimilar, a new framework may be required, perhaps accepting the perspective of a task-driven use of the visual cues, modulated by experience and cultural conditioning.     

Classification training in two- and three-year-old children

Two- and three-year-olds who did not group a set of geometric stimuli according to complete similarity on a pretest were divided into three experimental conditions: (a) a modeling condition in which the subjects observed a model group according to complete similarity while verbalizing his strategy, (b) a reinforcement condition in which the subjects were reinforced for sorting similar stimuli into boxes, and (c) a control condition. After the training sessions the Ss were given two posttests—one with stimuli identical to those used during the training session and one with a different set of geometric stimuli. On the posttests significantly more similarity classifications were obtained in the modeling condition than in either the reinforcement or the control conditions, which did not differ. It was concluded (a) that two- and three-year-olds can learn to group according to similarity, and (b) that modeling with verbalization is a very effective method for teaching young children to classify according to similarity.     

The implicit association test as a general measure of similarity.

The Implicit Association Test (IAT) is widely used as a measure of semantic similarity (i.e., associations in semantic memory). The results of previous research and of a new study show that IAT effects can, however, also be based on other types of similarity between stimuli. We therefore put forward the hypothesis that the IAT provides a general measure of similarity. Given that similarity is highly dynamic and context-dependent, our view that the IAT measures similarity is compatible with existing evidence showing that IAT effects are highly malleable. We provide further evidence for this in a new study in which the outcome of an IAT depended on whether the perceptual or functional characteristics of the stimuli were made salient.     

Comparison of the geometric and the contrast models of similarity by presentation of visual stimuli to the left and the right visual fields

In this study we investigated by means of the "same-different" decision task the process of comparing visual stimuli (schematic faces, familiar objects, houseplants, and nonsense figures) when presented for 100-150 msec to the right or to the left visual hemifields. The analysis of incorrect "same" responses showed that the addition of a common component (e.g., glasses, buttons) to a pair of nonidentical stimuli increased the percentage of incorrect same responses whereas the addition of the same component to one stimulus only in the pair decreased the percentage of incorrect "same" responses. This pattern, which is in accordance with Tversky's contrast model of similarity, is incompatible with any geometric model. Second, for schematic faces the results revealed that the left hemisphere is more sensitive to common than to distinctive features, whereas the right hemisphere is more sensitive to distinctive than to common features. No such interaction was obtained for the other type of stimuli. The implications of these results for models of similarity and the difference between the present findings and the findings of Sergent (1984) are discussed.     

知觉或记忆条件下对连续变化刺激的相似判断

控制不同材料(图形和字母)的大小、位置和长宽比连续变化,在标准和待判断刺激同时(知觉任务)或相继呈现(记忆任务)两种条件下进行相似判断的实验研究。结论为(1)任务方式对图形的相似判断有非常显著的影响;(2)知觉任务中相似判断的反应时非常显著地长于记忆任务中的反应时;(3)对材料各种特征的相似判断和加工速度差异非常显著．某些特征在相似表征中有较重要的作用。     

Processes of similarity judgment

Similarity underlies fundamental cognitive capabilities such as memory, categorization, decision making, problem solving, and reasoning. Although recent approaches to similarity appreciate the structure of mental representations, they differ in the processes posited to operate over these representations. We present an experiment that differentiates among extant structural accounts of similarity in their ability to account for patterns of similarity ratings. These data pose a challenge for transformation-based models and all but one mapping-based model, the Similarity as Interactive Activation and Mapping (SIAM) model of similarity.     

Similarity chains in the transformational paradigm

Rips’ (1989) results with the transformational paradigm have often been cited as supporting accounts of categorisation not based on similarity, such as involving necessary or sufficient features (or a belief in such features), which guarantee a categorisation outcome once their presence has been established. We discuss a similarity account of the transformational paradigm based on similarity chains, which predicts that when the transformation is more gradual the identity of the transformed object is less likely to change. Conversely, we suggest that an essentialist approach to categorisation predicts that essences are more likely to change in gradual transformations, across generations, as is the case with evolutionary change of species. In two experiments we examined the scope of the similarity versus the essentialist account in the transformational paradigm. With space aliens, the similarity account was superior to the essentialist one, but the converse was true with earth creatures. We suggest that an essentialist mode of categorisation is more likely than a similarity one for stimuli that are in better correspondence with our naïve understanding of the world.     

Similarity, kernels, and the triangle inequality

Similarity is used as an explanatory construct throughout psychology and multidimensional scaling (MDS) is the most popular way to assess similarity. In MDS, similarity is intimately connected to the idea of a geometric representation of stimuli in a perceptual space. Whilst connecting similarity and closeness of stimuli in a geometric representation may be intuitively plausible, Tversky and Gati [Tversky, A., & Gati, I. (1982). Similarity, separability, and the triangle inequality. Psychological Review, 89(2), 123-154] have reported data which are inconsistent with the usual geometric representations that are based on segmental additivity. We show that similarity measures based on Shepard’s universal law of generalization [Shepard, R. N. (1987). Toward a universal law of generalization for psychologica science. Science, 237(4820), 1317-1323] lead to an inner product representation in a reproducing kernel Hilbert space. In such a space stimuli are represented by their similarity to all other stimuli. This representation, based on Shepard’s law, has a natural metric that does not have additive segments whilst still retaining the intuitive notion of connecting similarity and distance between stimuli. Furthermore, this representation has the psychologically appealing property that the distance between stimuli is bounded.     

The planning and control model (PCM) of motorvisual priming: reconciling motorvisual impairment and facilitation effects

Previous research on dual-tasks has shown that, under some circumstances, actions impair the perception of action-consistent stimuli, whereas, under other conditions, actions facilitate the perception of action-consistent stimuli. We propose a new model to reconcile these contrasting findings. The planning and control model (PCM) of motorvisual priming proposes that action planning binds categorical representations of action features so that their availability for perceptual processing is inhibited. Thus, the perception of categorically action-consistent stimuli is impaired during action planning. Movement control processes, on the other hand, integrate multi-sensory spatial information about the movement and, therefore, facilitate perceptual processing of spatially movement-consistent stimuli. We show that the PCM is consistent with a wider range of empirical data than previous models on motorvisual priming. Furthermore, the model yields previously untested empirical predictions. We also discuss how the PCM relates to motorvisual research paradigms other than dual-tasks.     

Similarity effects in visual working memory

Perceptual similarity is an important property of multiple stimuli. Its computation supports a wide range of cognitive functions, including reasoning, categorization, and memory recognition. It is important, therefore, to determine why previous research has found conflicting effects of inter-item similarity on visual working memory. Studies reporting a similarity advantage have used simple stimuli whose similarity varied along a featural continuum. Studies reporting a similarity disadvantage have used complex stimuli from either a single or multiple categories. To elucidate stimulus conditions for similarity effects in visual working memory, we tested memory for complex stimuli (faces) whose similarity varied along a morph continuum. Participants encoded 3 morphs generated from a single face identity in the similar condition, or 3 morphs generated from different face identities in the dissimilar condition. After a brief delay, a test face appeared at one of the encoding locations for participants to make a same/different judgment. Two experiments showed that similarity enhanced memory accuracy without changing the response criterion. These findings support previous computational models that incorporate featural variance as a component of working memory load. They delineate limitations of models that emphasize cortical resources or response decisions.