On the Interaction of Theory and Data in Concept Learning

Standard models of concept learning generally focus on deriving statistical properties of a category based on data (i.e., category members and the features that describe them) but fail to give appropriate weight to the contact between people's intuitive theories and these data. Two experiments explored the role of people's prior knowledge or intuitive theories on category learning by manipulating the labels associated with the category. Learning differed dramatically when categories of children's drawings were meaningfully labeled (e.g., “done by creative children”) compared to when they were labeled in a neutral manner. When categories are meaningfully labeled, people bring intuitive theories to the learning context. Learning then involves a process in which people search for evidence in the data that supports abstract features or hypotheses that have been activated by the intuitive theories. In contrast, when categories are labeled in a neutral manner, people search for simple features that distinguish one category from another. Importantly, the final study suggests that learning involves an interaction of people's intuitive theories with data, in which theories and data mutually influence each other. The results strongly suggest that straight-forward, relatively modular ways of incorporating prior knowledge into models of category learning are inadequate. More telling, the results suggest that standard models may have fundamental limitations. We outline a speculative model of learning in which the interaction of theory and data is tightly coupled. The article concludes by comparing the results to recent artificial intelligence systems that use prior knowledge during learning.     

Incremental Bayesian Category Learning From Natural Language

Models of category learning have been extensively studied in cognitive science and primarily tested on perceptual abstractions or artificial stimuli. In this paper, we focus on categories acquired from natural language stimuli, that is, words (e.g., chair is a member of the furniture category). We present a Bayesian model that, unlike previous work, learns both categories and their features in a single process. We model category induction as two interrelated subproblems: (a) the acquisition of features that discriminate among categories, and (b) the grouping of concepts into categories based on those features. Our model learns categories incrementally using particle filters, a sequential Monte Carlo method commonly used for approximate probabilistic inference that sequentially integrates newly observed data and can be viewed as a plausible mechanism for human learning. Experimental results show that our incremental learner obtains meaningful categories which yield a closer fit to behavioral data compared to related models while at the same time acquiring features which characterize the learned categories. (An earlier version of this work was published in Frermann and Lapata 2014 .)     

Order Matters! Influences of Linear Order on Linguistic Category Learning

Linguistic category learning has been shown to be highly sensitive to linear order, and depending on the task, differentially sensitive to the information provided by preceding category markers (premarkers, e.g., gendered articles) or succeeding category markers (postmarkers, e.g., gendered suffixes). Given that numerous systems for marking grammatical categories exist in natural languages, it follows that a better understanding of these findings can shed light on the factors underlying this diversity. In two discriminative learning simulations and an artificial language learning experiment, we identify two factors that modulate linear order effects in linguistic category learning: category structure and the level of abstraction in a category hierarchy. Regarding category structure, we find that postmarking brings an advantage for learning category diagnostic stimulus dimensions, an effect not present when categories are non-confusable. Regarding levels of abstraction, we find that premarking of super-ordinate categories (e.g., noun class) facilitates learning of subordinate categories (e.g., nouns). We present detailed simulations using a plausible candidate mechanism for the observed effects, along with a comprehensive analysis of linear order effects within an expectation-based account of learning. Our findings indicate that linguistic category learning is differentially guided by pre- and postmarking, and that the influence of each is modulated by the specific characteristics of a given category system.     

Visual variability affects early verb learning

Research demonstrates that within‐category visual variability facilitates noun learning; however, the effect of visual variability on verb learning is unknown. We habituated 24‐month‐old children to a novel verb paired with an animated star‐shaped actor. Across multiple trials, children saw either a single action from an action category (identical actions condition, for example, travelling while repeatedly changing into a circle shape) or multiple actions from that action category (variable actions condition, for example, travelling while changing into a circle shape, then a square shape, then a triangle shape). Four test trials followed habituation. One paired the habituated verb with a new action from the habituated category (e.g., ‘dacking’ + pentagon shape) and one with a completely novel action (e.g., ‘dacking’ + leg movement). The others paired a new verb with a new same‐category action (e.g., ‘keefing’ + pentagon shape), or a completely novel category action (e.g., ‘keefing’ + leg movement). Although all children discriminated novel verb/action pairs, children in the identical actions condition discriminated trials that included the completely novel verb, while children in the variable actions condition discriminated the out‐of‐category action. These data suggest that – as in noun learning – visual variability affects verb learning and children's ability to form action categories.     

Speaker variability augments phonological processing in early word learning

Infants in the early stages of word learning have difficulty learning lexical neighbors (i.e. word pairs that differ by a single phoneme), despite their ability to discriminate the same contrast in a purely auditory task. While prior work has focused on top‐down explanations for this failure (e.g. task demands, lexical competition), none has examined if bottom‐up acoustic‐phonetic factors play a role. We hypothesized that lexical neighbor learning could be improved by incorporating greater acoustic variability in the words being learned, as this may buttress still‐developing phonetic categories, and help infants identify the relevant contrastive dimension. Infants were exposed to pictures accompanied by labels spoken by either a single or multiple speakers. At test, infants in the single‐speaker condition failed to recognize the difference between the two words, while infants who heard multiple speakers discriminated between them.     

Does Infant-Directed Speech Help Phonetic Learning? A Machine Learning Investigation

A prominent hypothesis holds that by speaking to infants in infant-directed speech (IDS) as opposed to adult-directed speech (ADS), parents help them learn phonetic categories. Specifically, two characteristics of IDS have been claimed to facilitate learning: hyperarticulation, which makes the categories more separable, and variability, which makes the generalization more robust. Here, we test the separability and robustness of vowel category learning on acoustic representations of speech uttered by Japanese adults in ADS, IDS (addressed to 18- to 24-month olds), or read speech (RS). Separability is determined by means of a distance measure computed between the five short vowel categories of Japanese, while robustness is assessed by testing the ability of six different machine learning algorithms trained to classify vowels to generalize on stimuli spoken by a novel speaker in ADS. Using two different speech representations, we find that hyperarticulated speech, in the case of RS, can yield better separability, and that increased between-speaker variability in ADS can yield, for some algorithms, more robust categories. However, these conclusions do not apply to IDS, which turned out to yield neither more separable nor more robust categories compared to ADS inputs. We discuss the usefulness of machine learning algorithms run on real data to test hypotheses about the functional role of IDS.     

Inhibition of irrelevant category–response mappings

When switching tasks, performance tends to be worse for n – 2 repetitions than with n – 2 switches. This n – 2 repetition cost has been hypothesized to reflect task-set inhibition: specifically, inhibition of irrelevant category–response mappings involved in response selection. This hypothesis leads to divergent predictions for situations in which all tasks involve the same stimulus categories: An n – 2 repetition cost is predicted when response sets differ across tasks, but not when the response set stays the same. The authors tested these predictions by having subjects perform relative judgements with different reference points. In Experiment 1, the stimulus categories were the same across reference points, but the response set either differed or stayed the same (the multiple- and single-mapping conditions, respectively). An n – 2 repetition cost was found in the multiple-mapping condition but not in the single-mapping condition. Experiment 2 provided evidence against the possibility that these divergent effects reflected differences in memory load. These findings confirm predictions that link n – 2 repetition costs to inhibition of irrelevant category–response mappings.     

Conceptual complexity and the bias/variance tradeoff

In this paper we propose that the conventional dichotomy between exemplar-based and prototype-based models of concept learning is helpfully viewed as an instance of what is known in the statistical learning literature as the bias/variance tradeoff. The bias/variance tradeoff can be thought of as a sliding scale that modulates how closely any learning procedure adheres to its training data. At one end of the scale (high variance), models can entertain very complex hypotheses, allowing them to fit a wide variety of data very closely—but as a result can generalize poorly, a phenomenon called overfitting. At the other end of the scale (high bias), models make relatively simple and inflexible assumptions, and as a result may fit the data poorly, called underfitting. Exemplar and prototype models of category formation are at opposite ends of this scale: prototype models are highly biased, in that they assume a simple, standard conceptual form (the prototype), while exemplar models have very little bias but high variance, allowing them to fit virtually any combination of training data. We investigated human learners’ position on this spectrum by confronting them with category structures at variable levels of intrinsic complexity, ranging from simple prototype-like categories to much more complex multimodal ones. The results show that human learners adopt an intermediate point on the bias/variance continuum, inconsistent with either of the poles occupied by most conventional approaches. We present a simple model that adjusts (regularizes) the complexity of its hypotheses in order to suit the training data, which fits the experimental data better than representative exemplar and prototype models.     

The advantages of mathematical modeling over traditional methods in the analysis of category clustering

A mathematical model for the analysis of category clustering is developed and testd. The model, which can be applied to categories of any size, is an extension of a two-item statistical model developed by Batchelder and Riefer (Psychological Review, 1980, 87, 375–397), and is equivalent to their model when categories consist of two items. The model is based on a current theory of clustering which postulates that the learning of a list of category items occurs on different hierarchical levels. Two category list-learning experiments are presented, and the data from these experiments are analyzed using the general statistical model. The first experiment reveals that the probabilities of storing and retrieving a cluster increase with category size, while the learning of items as singletons decreases. The effects of within-category spacing indicate that the storage of clusters decreases while cluster retrievability increases with an increase in input spacing. In the second experiment, the storage and retrieval of clusters are shown to be unaffected by whether the presentation of items is uncued or cued with the name of the category. However, the association of items decreases and the learning of items as singletons increases with uncued presentation. In the final sections, the general statistical model is compared to other methods for the measurement of category clustering. The model is shown to be superior to numerical indices of clustering, since these measures are not based on any theory of clustering, and because unitary measures cannot capture the multiprocess nature of categorized recall. The model is also argued to have certain advantages over other mathematical models that have been applied to category clustering, since these models cannot account for situations in which a portion of the items are clustered while others are learned singularly.     

Children learn words easier when they are interested in the category to which the word belongs

The overall pattern of vocabulary development is relatively similar across children learning different languages. However, there are considerable differences in the words known to individual children. Historically, this variability has been explained in terms of differences in the input. Here, we examine the alternate possibility that children's individual interest in specific natural categories shapes the words they are likely to learn – a child who is more interested in animals will learn a new animal name easier relative to a new vehicle name. Two‐year‐old German‐learning children (N = 39) were exposed to four novel word–object associations for objects from four different categories. Prior to the word learning task, we measured their interest in the categories that the objects belonged to. Our measure was pupillary change following exposure to familiar objects from these four categories, with increased pupillary change interpreted as increased interest in that category. Children showed more robust learning of word–object associations from categories they were more interested in relative to categories they were less interested in. We further found that interest in the novel objects themselves influenced learning, with distinct influences of both category interest and object interest on learning. These results suggest that children's interest in different natural categories shapes their word learning. This provides evidence for the strikingly intuitive possibility that a child who is more interested in animals will learn novel animal names easier than a child who is more interested in vehicles.     

The impact of training sequence and between-category similarity on unsupervised induction

Studies of supervised categorization often show better learning when examples are presented in random alternation rather than massed by category, but such interleaving impairs learning in unsupervised tasks. The exemplar comparison hypothesis explains this result by assuming that people in unsupervised tasks discover generalizations about categories by comparing individual examples, and that interleaving increases the difficulty of such within-category comparisons. The category invention hypothesis explains the interleaving effect by assuming that people are more likely to merge or aggregate potentially separable categories when they are interleaved, and this initial failure to recognize separate categories then acts as an effective barrier to further learning. The present experiments show that the interleaving effect depends on the similarity or alignability of the presented categories. This result provides evidence in favour of the category invention hypothesis, which expects that highly dissimilar (nonalignable) categories will resist aggregation and hence will not be affected by interleaving. The nonmonotonic pattern of learning, and the interaction between sequence and similarity, observed in the alignable conditions of Experiment 3 were also consistent with category invention, but not with exemplar comparison. Implications are discussed for real-world learning, especially the relationship between exposure and learning and between supervised and unsupervised learning.     

EARSHOT: A Minimal Neural Network Model of Incremental Human Speech Recognition

Despite the lack of invariance problem (the many-to-many mapping between acoustics and percepts), human listeners experience phonetic constancy and typically perceive what a speaker intends. Most models of human speech recognition (HSR) have side-stepped this problem, working with abstract, idealized inputs and deferring the challenge of working with real speech. In contrast, carefully engineered deep learning networks allow robust, real-world automatic speech recognition (ASR). However, the complexities of deep learning architectures and training regimens make it difficult to use them to provide direct insights into mechanisms that may support HSR. In this brief article, we report preliminary results from a two-layer network that borrows one element from ASR, long short-term memory nodes, which provide dynamic memory for a range of temporal spans. This allows the model to learn to map real speech from multiple talkers to semantic targets with high accuracy, with human-like timecourse of lexical access and phonological competition. Internal representations emerge that resemble phonetically organized responses in human superior temporal gyrus, suggesting that the model develops a distributed phonological code despite no explicit training on phonetic or phonemic targets. The ability to work with real speech is a major advance for cognitive models of HSR.     

Learning words’ sounds before learning how words sound: 9-Month-olds use distinct objects as cues to categorize speech information

One of the central themes in the study of language acquisition is the gap between the linguistic knowledge that learners demonstrate, and the apparent inadequacy of linguistic input to support induction of this knowledge. One of the first linguistic abilities in the course of development to exemplify this problem is in speech perception: specifically, learning the sound system of one’s native language. Native-language sound systems are defined by meaningful contrasts among words in a language, yet infants learn these sound patterns before any significant numbers of words are acquired. Previous approaches to this learning problem have suggested that infants can learn phonetic categories from statistical analysis of auditory input, without regard to word referents. Experimental evidence presented here suggests instead that young infants can use visual cues present in word-labeling situations to categorize phonetic information. In Experiment 1, 9-month-old English-learning infants failed to discriminate two non-native phonetic categories, establishing baseline performance in a perceptual discrimination task. In Experiment 2, these infants succeeded at discrimination after watching contrasting visual cues (i.e., videos of two novel objects) paired consistently with the two non-native phonetic categories. In Experiment 3, these infants failed at discrimination after watching the same visual cues, but paired inconsistently with the two phonetic categories. At an age before which memory of word labels is demonstrated in the laboratory, 9-month-old infants use contrastive pairings between objects and sounds to influence their phonetic sensitivity. Phonetic learning may have a more functional basis than previous statistical learning mechanisms assume: infants may use cross-modal associations inherent in social contexts to learn native-language phonetic categories.     

Dual-learning systems during speech category learning

Dual-system models of visual category learning posit the existence of an explicit, hypothesis-testing reflective system, as well as an implicit, procedural-based reflexive system. The reflective and reflexive learning systems are competitive and neurally dissociable. Relatively little is known about the role of these domain-general learning systems in speech category learning. Given the multidimensional, redundant, and variable nature of acoustic cues in speech categories, our working hypothesis is that speech categories are learned reflexively. To this end, we examined the relative contribution of these learning systems to speech learning in adults. Native English speakers learned to categorize Mandarin tone categories over 480 trials. The training protocol involved trial-by-trial feedback and multiple talkers. Experiments 1 and 2 examined the effect of manipulating the timing (immediate vs. delayed) and information content (full vs. minimal) of feedback. Dual-system models of visual category learning predict that delayed feedback and providing rich, informational feedback enhance reflective learning, while immediate and minimally informative feedback enhance reflexive learning. Across the two experiments, our results show that feedback manipulations that targeted reflexive learning enhanced category learning success. In Experiment 3, we examined the role of trial-to-trial talker information (mixed vs. blocked presentation) on speech category learning success. We hypothesized that the mixed condition would enhance reflexive learning by not allowing an association between talker-related acoustic cues and speech categories. Our results show that the mixed talker condition led to relatively greater accuracies. Our experiments demonstrate that speech categories are optimally learned by training methods that target the reflexive learning system.     

Inferior Frontal Regions Underlie the Perception of Phonetic Category Invariance

ABSTRACT— The problem of mapping differing sensory stimuli onto a common category is fundamental to human cognition. Listeners perceive stable phonetic categories despite many sources of acoustic variability. What are the neural mechanisms that underlie this perceptual stability? In this functional magnetic resonance imaging study, a short-interval habituation paradigm was used to investigate neural sensitivity to acoustic changes within and between phonetic categories. A region in the left inferior frontal sulcus showed a pattern of activation consistent with phonetic invariance: insensitivity to acoustic changes within a phonetic category and sensitivity to changes between phonetic categories. Left superior temporal regions, in contrast, showed graded sensitivity to both within- and between-category changes. These results suggest that perceptual insensitivity to changes within a phonetic category may arise from decision-related mechanisms in the left prefrontal cortex and add to a growing body of literature suggesting that the inferior prefrontal cortex plays a domain-general role in computing category representations.     

Alternating Incubation Effects in the Generation of Category Exemplars

Four experiments tested the forgetting fixation hypothesis of incubation effects, comparing continuous vs. alternating generation of exemplars from three different types of categories. In two experiments, participants who listed as many members as possible from two different categories produced more responses, and more novel responses, when they alternated back and forth between the two categories, as compared to continuous uninterrupted listing from each of the two categories. This incubation effect was not found in Experiment 1, when participants were given taxonomic categories (birds and clothing ) for the generation task, but was found in Experiment 2 with sense impression categories (cold things and heavy things ), and in Experiment 3 with ad hoc categories (equipment you take camping and fattening foods ). A similar incubation effect was observed in Experiment 4 when a non‐verbal task was given between category generation tasks, but only for flexibly defined categories. The results suggest that forgetting from one alternating listing period to the next in the form of altering category cue representations was consistent with the observed incubation effects. These alternating incubation effects have implications for understanding cognitive processes that underlie creative cognition.     

Neural markers of categorization in 6-month-old infants

Little is known of the neural processes that underlie concept-formation abilities in human infants. We investigated category-learning processes in infants both by using a common behavioral measure and by recording the brain's electrical activity (event-related potentials, or ERPs). ERPs were recorded while 6-month-olds viewed cat images during training, followed by novel cat images interspersed with novel dog images during test. The data indicate that distinct neural signals correspond with learning of a category presented during familiarization, preferential responding to a novel category, and representation of category exemplars at multiple levels of inclusiveness. The results suggest that fundamental components of the neural architecture supporting object categorization are functional within the first half-year of postnatal life, before infants acquire language and young children engage in formal learning of semantic categories. The findings are discussed in terms of their implications for models of category learning and development.     

类别学习的阻碍效应与双机制理论

通过对类别学习中的阻碍效应进行系统研究,尝试性的提出了类别学习的双机制理论。三个实验分别考察：当样例特征随机呈现时;定义特征固定呈现在样例首位时;以及刺激材料为知觉图形材料时,类别学习中的阻碍效应。三个实验的研究结果都发现：在类别学习中存在一定程度的阻碍效应,支持类别学习同时存在联结学习机制和认知学习机制的双机制观点。     

Category length produces an inverted-U discriminability function in episodic recognition memory

Four experiments used signal detection analyses to assess recognition memory for lists of words consisting of differing numbers of exemplars from different semantic categories. The results showed that recognition memory performance, measured by d _a , (a) increased as category length (CL, the number of study-list items selected from the same semantic category) increased from 1 to 8 but then decreased as CL further increased from 8 to 14, (b) was greater when 2 studied items from the same category occurred back to back, rather than being separated by 5–11 items from other categories, and (c) was greater for the first studied 2 exemplars than for the last studied 2 exemplars from blocked categories with CLs of 8 or 14. For all CLs, all z-coordinate receiver operating characteristic (z-ROC) functions were linear with slopes significantly less than 1.0, and none had a significant quadratic component. These results pose a challenge for three major classes of recognition memory models: item-noise, context-noise, and dual-process models.     

Auditory and phonetic coding of the cues for speech: Discrimination of the [r-l] distinction by young infants

Infants, 2 and 3 months of age, were found to discriminte stimuli along the acoustic continuum underlying the phonetic contrast [r] vs. [l] in a nearly categorical manner. For an approximately equal acoustic difference, discrimination, as measured by recovery from satiation or familiarization, was reliably better when the two stimuli were exemplars of different phonetic categories than when they were acoustic variations of the same phonetic category. Discrimination of the same acoustic information presented in a nonspeech mode was found to be continuous, that is, determined by acoustic rather than phonetic characteristics of the stimuli. The findings were discussed with reference to the nature of the mechanisms that may determine the processing of complex acoustic signals in young infants and with reference to the role of linguistic experience on the development of speech perception at the phonetic level.