Statistical phonetic learning in infants: facilitation and feature generalization

Over the course of the first year of life, infants develop from being generalized listeners, capable of discriminating both native and non-native speech contrasts, into specialized listeners whose discrimination patterns closely reflect the phonetic system of the native language(s). Recent work by Maye, Werker and Gerken (2002) has proposed a statistical account for this phenomenon, showing that infants may lose the ability to discriminate some foreign language contrasts on the basis of their sensitivity to the statistical distribution of sounds in the input language. In this paper we examine the process of enhancement in infant speech perception, whereby initially difficult phonetic contrasts become better discriminated when they define two categories that serve a functional role in the native language. In particular, we demonstrate that exposure to a bimodal statistical distribution in 8-month-old infants' phonetic input can lead to increased discrimination of difficult contrasts. In addition, this exposure also facilitates discrimination of an unfamiliar contrast sharing the same phonetic feature as the contrast presented during familiarization, suggesting that infants extract acoustic/phonetic information that is invariant across an abstract featural representation.     

On the internal structure of phonetic categories: a progress report

There is growing evidence that phonetic categories have a rich internal structure, with category members varying systematically in category goodness. Our recent findings on this issue, which are summarized in this paper, underscore the existence and robustness of this structure and indicate further that the mapping between acoustic signal and internal category structure is complex: just as in the case of category boundaries, the best exemplars of a given category are highly dependent on acoustic-phonetic context and are specified by multiple properties of the speech signal. These findings suggest that the listener's representation of phonetic form preserves not only vategorical information, but also fine-grained information about the detailed acoustic-phonetic characteristics of the language.     

Hippocampus, cortex, and basal ganglia: insights from computational models of complementary learning systems

We present a framework for understanding how the hippocampus, neocortex, and basal ganglia work together to support cognitive and behavioral function in the mammalian brain. This framework is based on computational tradeoffs that arise in neural network models, where achieving one type of learning function requires very different parameters from those necessary to achieve another form of learning. For example, we dissociate the hippocampus from cortex with respect to general levels of activity, learning rate, and level of overlap between activation patterns. Similarly, the frontal cortex and associated basal ganglia system have important neural specializations not required of the posterior cortex system. Taken together, this overall cognitive architecture, which has been implemented in functioning computational models, provides a rich and often subtle means of explaining a wide range of behavioral and cognitive neuroscience data. Here, we summarize recent results in the domains of recognition memory, contextual fear conditioning, effects of basal ganglia lesions on stimulus-response and place learning, and flexible responding.     

Visual influences on the internal structure of phonetic categories

Previous work has demonstrated that the graded internal structure of phonetic categories is sensitive to a variety of contextual factors. One such factor is place of articulation: The best exemplars of voiceless stop consonants along auditory bilabial and velar voice onset time (VOT) continua occur over different ranges of VOTs (Volaitis & Miller, 1992). In the present study, we exploited the McGurk effect to examine whether visual information for place of articulation also shifts the best exemplar range for voiceless consonants, following Green and Kuhl's (1989) demonstration of effects of visual place of articulation on the location of voicing boundaries. In Experiment 1, we established that /p/ and /t/ have different best exemplar ranges along auditory bilabial and alveolar VOT continua. We then found, in Experiment 2, a similar shift in the best-exemplar range for /t/ relative to that for /p/ when there was a change in visual place of articulation, with auditory place of articulation held constant. These findings indicate that the perceptual mechanisms that determine internal phonetic category structure are sensitive to visual, as well as to auditory, information.     

Selective attention and the acquisition of new phonetic categories

A class of selective attention models often applied to speech perception is used to study effects of training on the perception of an unfamiliar phonetic contrast. Attention-to-dimension (A2D) models of perceptual learning assume that the dimensions that structure listeners' perceptual space are constant and that learning involves only the reweighting of existing dimensions to emphasize or de-emphasize different sensory dimensions. Multidimensional scaling is used to identify the acoustic-phonetic dimensions listeners use before and after training to recognize the 3 classes of Korean stop consonants. Results suggest that A2D models can account for some observed restructuring of listeners' perceptual space, but listeners also show evidence of directing attention to a previously unattended dimension of phonetic contrast.     

Speech-sound duration processing in a second language is specific to phonetic categories

The mismatch negativity (MMN) component of the auditory event-related potential was used to determine the effect of native language, Russian, on the processing of speech-sound duration in a second language, Finnish, that uses duration as a cue for phonological distinction. The native-language effect was compared with Finnish vowels that either can or cannot be categorized using the Russian phonological system. The results showed that the duration-change MMN for the Finnish sounds that could be categorized through Russian was reduced in comparison with that for the Finnish sounds having no Russian equivalent. In the Finnish sounds that can be mapped through the Russian phonological system, the facilitation of the duration processing may be inhibited by the native Russian language. However, for the sounds that have no Russian equivalent, new vowel categories independent of the native Russian language have apparently been established, enabling a native-like duration processing of Finnish.     

Recalibration of phonetic categories by lipread speech versus lexical information

Listeners hearing an ambiguous phoneme flexibly adjust their phonetic categories in accordance with information telling what the phoneme should be (i.e., recalibration). Here the authors compared recalibration induced by lipread versus lexical information. Listeners were exposed to an ambiguous phoneme halfway between /t/ and /p/ dubbed onto a face articulating /t/ or /p/ or embedded in a Dutch word ending in /t/ (e.g., groot [big]) or /p/ (knoop [button]). In a posttest, participants then categorized auditory tokens as /t/ or /p/. Lipread and lexical aftereffects were comparable in size (Experiment 1), dissipated about equally fast (Experiment 2), were enhanced by exposure to a contrast phoneme (Experiment 3), and were not affected by a 3-min silence interval (Experiment 4). Exposing participants to 1 instead of both phoneme categories did not make the phenomenon more robust (Experiment 5). Despite the difference in nature (bottom-up vs. top-down information), lipread and lexical information thus appear to serve a similar role in phonetic adjustments.     

Cerebral hemispheric involvement in the acquisition of new phonetic categories

This study was concerned with the involvement of the cerebral hemispheres in the acquisition of perceptual skill with novel speech sounds. Two groups of eight subjects rated on three occasions the dissimilarity of pairs of Mandarin Chinese speech sounds varying on voicing and aspiration, presented to the left or right ear with contralateral noise. The experimental group received listening experience with long passages of Mandarin containing the target sounds. Multidimensional scaling analysis of dissimilarity ratings indicated that listening experience leads to increased perceptual differentiation of phonetic categories drawn from a language unfamiliar to the listener. This improvement occurred sooner with presentation of the target sounds to the right hemisphere than to the left if the phonological contrast was along the voicing dimension. Improvement in the perception of stimuli varying on aspiration occurred in right ear presentations only. This finding supports the position that speech perception mechanisms at the feature level may be distributed asymmetrically across the hemispheres.     

The acquisition of phonetic categories in bilingual infants: new data from an anticipatory eye movement paradigm

Contrasting results have been reported regarding the phonetic acquisition of bilinguals. A lack of discrimination has been observed for certain native contrasts in 8-month-old Catalan-Spanish bilingual infants (Bosch & Sebastián-Gallés, 2003a), though not in French-English bilingual infants (Burns, Yoshida, Hill & Werker, 2007; Sundara, Polka & Molnar, 2008). At present, the data for Catalan-Spanish bilinguals constitute an exception in the early language acquisition literature. This study contributes new findings that show that Catalan-Spanish bilingual infants do not lose the capacity to discriminate native contrasts. We used an adaptation of the anticipatory eye movement paradigm (AEM; McMurray & Aslin, 2004) to explore this question. In two experiments we tested the ability of infants from Catalan and Spanish monolingual families and from Catalan-Spanish bilingual families to discriminate a Spanish-Catalan common and a Catalan-specific vowel contrast. Results from both experiments revealed that Catalan-Spanish bilingual infants showed the same discrimination abilities as those shown by their monolingual peers, even in a phonetic contrast that had not been discriminated in previous studies. Our results demonstrate that discrimination can be observed in 8-month-old bilingual infants when tested with a measure not based on recovery of attention. The high ratio of cognates in Spanish and Catalan may underlie the reason why bilinguals failed to discriminate the native vowels when tested with the familiarization-preference procedure but succeeded with the AEM paradigm.     

Procedural learning of unstructured categories

Unstructured categories are those in which the stimuli are assigned to each contrasting category randomly, and thus there is no rule- or similarity-based strategy for determining category membership. Intuition suggests that unstructured categories are likely to be learned via explicit memorization that is under the control of declarative memory. In contrast to this prediction, neuroimaging studies of unstructured-category learning have reported task-related activation in the striatum, but typically not in the hippocampus—results that seem more consistent with procedural learning than with a declarative-memory strategy. This article reports the first known behavioral test of whether unstructured-category learning is mediated by explicit strategies or by procedural learning. Our results suggest that the feedback-based learning of unstructured categories is mediated by procedural memory.     

Temporal constraints on visual learning: a computational model

Given a constant stream of perceptual stimuli, how can the underlying invariances associated with a given input be learned? One approach consists of using generic truths about the spatiotemporal structure of the physical world as constraints on the types of quantities learned. The learning methodology employed here embodies one such truth: that perceptually salient properties (such as stereo disparity) tend to vary smoothly over time. Unfortunately, the units of an artificial neural network tend to encode superficial image properties, such as individual grey-level pixel values, which vary rapidly over time. However, if the states of units are constrained to vary slowly, then the network is forced to learn a smoothly varying function of the training data. We implemented this temporal-smoothness constraint in a backpropagation network which learned stereo disparity from random-dot stereograms. Temporal smoothness was formalized with the use of regularization theory by modifying the standard cost function minimised during training of a network. Temporal smoothness was found to be similar to other techniques for improving generalisation, such as early stopping and weight decay. However, in contrast to these, the theoretical underpinnings of temporal smoothing are intimately related to fundamental characteristics of the physical world. Results are discussed in terms of regularization theory and the physically realistic assumptions upon which temporal smoothing is based.     

Idealness and similarity in goal-derived categories: A computational examination

The finding that the typicality gradient in goal-derived categories is mainly driven by ideals rather than by exemplar similarity has stood uncontested for nearly three decades. Due to the rather rigid earlier implementations of similarity, a key question has remained—that is, whether a more flexible approach to similarity would alter the conclusions. In the present study, we evaluated whether a similarity-based approach that allows for dimensional weighting could account for findings in goal-derived categories. To this end, we compared a computational model of exemplar similarity (the generalized context model; Nosofsky, Journal of Experimental Psychology. General 115:39–57, 1986) and a computational model of ideal representation (the ideal-dimension model; Voorspoels, Vanpaemel, & Storms, Psychonomic Bulletin & Review 18:1006-114, 2011) in their accounts of exemplar typicality in ten goal-derived categories. In terms of both goodness-of-fit and generalizability, we found strong evidence for an ideal approach in nearly all categories. We conclude that focusing on a limited set of features is necessary but not sufficient to account for the observed typicality gradient. A second aspect of ideal representations—that is, that extreme rather than common, central-tendency values drive typicality—seems to be crucial.     

Contextual influences on the internal structure of phonetic categories: a distinction between lexical status and speaking rate

Previous research has shown that phonetic categories have a graded internal structure that is highly dependent on acoustic-phonetic contextual factors, such as speaking rate; these factors alter not only the location of phonetic category boundaries, but also the location of a category's best exemplars. The purpose of the present investigation, which focused on the voiceless category as specified by voice onset time (VOT), was to determine whether a higher order linguistic contextual factor, lexical status, which is known to alter the location of the voiced-voiceless phonetic category boundary, also alters the location of the best exemplars of the voiceless category. The results indicated that lexical status has a more limited and qualitatively different effect on the category's best exemplars than does the acoustic-phonetic factor of speaking rate. This dissociation is discussed in terms of a production-based account in which perceived best exemplars of a category track contextual variation in speech production.     

Processing empty categories: A parallel approach

In many current theories of human sentence processing, the mechanisms and principles proposed to account for data related to processing (1) long-distance dependencies and (2) structural ambiguities, such as garden-path sentences, are independent, despite the fact that deciding whether or not to posit a gap is just a special case of ambiguity resolution. In this paper we demonstrate how the parallel parsing theory proposed by Gibson (1991, in press)—which was developed to account for nongap ambiguity resolution data—also explains a number of gap-positing facts, without additional strategies. In particular, we show how this theory correctly explains filled-object-gap effects and the lack of filled-subject-gap effects in English, as well as certain gap-processing effects in Dutch.     

Sorting out categories: Incremental learning of category structure

Two experiments examine how inferences might promote unsupervised and incremental category learning. Many categories have members related through overall similarity (e.g., a family resemblance structure) rather than by a defining feature. However, when people are asked to sort category members in a category construction task, they often do so by partitioning on a single feature. Starting from an earlier result showing that pairwise inferences increase family resemblance sorting (Lassaline & Murphy, 1996), we examine how these inferences lead to learning the family resemblance structure. Results show that the category structure is learned incrementally. The pairwise inferences influence participants’ weightings of feature pairs that were specifically asked about, which in turn affects their sorting. The sorting then allows further learning of the categorical structure. Thus, the inferences do not directly lead learners to the family resemblance structure, but they do provide a foundation to build on as the participants make additional judgments.     

Hemispheric asymmetry of auditory evoked potentials to comparisons within and across phonetic categories

In 10 right-handed Ss, auditory evoked responses (AERs) were recorded from left and right temporal and parietal scalp regions during simple discrimination responses to binaurally presented pairs of synthetic speech sounds ranging perceptually from /ba/ to /da/. A late positive component (P3) in the AER was found to reflect the categorical or phonetic analysis of the stop consonants, with only left scalp sites averaging significantly different responses between acoustic and phonetic comparisons. The result is interpreted as evidence of hemispheric differences in the processing of speech in respect of the level of processing accessed by the particular information processing task.     

Learning categories at different hierarchical levels: a comparison of category learning models

Three formal models of category learning, the rational model (Anderson, 1990), the configural-cue model (Gluck & Bower, 1988a), and ALCOVE (Kruschke, 1992), were evaluated on their ability to account for differential learning of hierarchically structured categories. An experiment using a theoretically challenging category structure developed by Lassaline, Wisniewski, and Medin (1992) is reported. Subjects learned one of two different category structures. For one structure, diagnostic information was present along a single dimension (1-D). For the other structure, diagnostic information was distributed across four dimensions (4-D). Subjects learned these categories at a general or at a specific level of abstraction. For the 1-D structure, specific-level categories were learned more rapidly than general-level categories. For the 4-D structure, the opposite result was observed. These results proved highly diagnostic for evaluating the models—although ALCOVE provided a good account of the observed results, the rational model and the configural-cue model did not.     

Grounding words in perception and action: computational insights

We use words to communicate about things and kinds of things, their properties, relations and actions. Researchers are now creating robotic and simulated systems that ground language in machine perception and action, mirroring human abilities. A new kind of computational model is emerging from this work that bridges the symbolic realm of language with the physical realm of real-world referents. It explains aspects of context-dependent shifts of word meaning that cannot easily be explained by purely symbolic models. An exciting implication for cognitive modeling is the use of grounded systems to 'step into the shoes' of humans by directly processing first-person-perspective sensory data, providing a new methodology for testing various hypotheses of situated communication and learning.     

Implicit learning of tonality: a self-organizing approach

Tonal music is a highly structured system that is ubiquitous in our cultural environment. We demonstrate the acquisition of implicit knowledge of tonal structure through neural self-organization resulting from mere exposure to simultaneous and sequential combinations of tones. In the process of learning, a network with fundamental neural constraints comes to internalize the essential correlational structure of tonal music. After learning, the network was run through a range of experiments from the literature. The model provides a parsimonious account of a variety of empirical findings dealing with the processing of tone, chord, and key relationships, including relatedness judgments, memory judgments, and expectancies. It also illustrates the plausibility of activation being a unifying mechanism underlying a range of cognitive tasks.     

Accounting for the computational basis of consciousness: a connectionist approach

This paper argues for an explanation of the mechanistic (computational) basis of consciousness that is based on the distinction between localist (symbolic) representation and distributed representation, the ideas of which have been put forth in the connectionist literature. A model is developed to substantiate and test this approach. The paper also explores the issue of the functional roles of consciousness, in relation to the proposed mechanistic explanation of consciousness. The model, embodying the representational difference, is able to account for the functional role of consciousness, in the form of the synergy between the conscious and the unconscious. The fit between the model and various cognitive phenomena and data (documented in the psychological literatures) is discussed to accentuate the plausibility of the model and its explanation of consciousness. Comparisons with existing models of consciousness are made in the end.