The Learning Signal in Perceptual Tuning of Speech: Bottom Up Versus Top-Down Information

Cognitive systems face a tension between stability and plasticity. The maintenance of long-term representations that reflect the global regularities of the environment is often at odds with pressure to flexibly adjust to short-term input regularities that may deviate from the norm. This tension is abundantly clear in speech communication when talkers with accents or dialects produce input that deviates from a listener's language community norms. Prior research demonstrates that when bottom-up acoustic information or top-down word knowledge is available to disambiguate speech input, there is short-term adaptive plasticity such that subsequent speech perception is shifted even in the absence of the disambiguating information. Although such effects are well-documented, it is not yet known whether bottom-up and top-down resolution of ambiguity may operate through common processes, or how these information sources may interact in guiding the adaptive plasticity of speech perception. The present study investigates the joint contributions of bottom-up information from the acoustic signal and top-down information from lexical knowledge in the adaptive plasticity of speech categorization according to short-term input regularities. The results implicate speech category activation, whether from top-down or bottom-up sources, in driving rapid adjustment of listeners' reliance on acoustic dimensions in speech categorization. Broadly, this pattern of perception is consistent with dynamic mapping of input to category representations that is flexibly tuned according to interactive processing accommodating both lexical knowledge and idiosyncrasies of the acoustic input.     

Repetition priming of words, pseudowords, and nonwords

In 5 experiments, the authors assessed repetition priming for words, pseudowords, and nonwords using a task that combines an implicit perceptual fluency measure and a recognition memory assessment for each list item. Words and pseudowords generated a consistently strong repetition effect even when there was a failure to recognize the stimulus. In 2 of the experiments, the repetition effect for nonwords was reliably above chance even when there was a failure to recognize the stimulus. The authors propose a parallel distributed processing (PDP) model based on the work of J. McClelland and D. Rumelhart (1985) as a way to understand the mechanisms potentially responsible for the pattern of findings. Although the error-driven nature of learning in the model results in a poor fit to the nonword priming data, this is not endemic to all PDP models. Using a model based on Hebbian learning, the authors instantiate a property that they believe is characteristic of implicit memory--that learning is primarily based on the strengthening of connections between units that become active during the processing of a stimulus. This model provides a far more satisfactory account of the data than does the error-driven model.     

ALCOVE: an exemplar-based connectionist model of category learning.

ALCOVE (attention learning covering map) is a connectionist model of category learning that incorporates an exemplar-based representation (Medin & Schaffer, 1978; Nosofsky, 1986) with error-driven learning (Gluck & Bower, 1988; Rumelhart, Hinton, & Williams, 1986). Alcove selectively attends to relevant stimulus dimensions, is sensitive to correlated dimensions, can account for a form of base-rate neglect, does not suffer catastrophic forgetting, and can exhibit 3-stage (U-shaped) learning of high-frequency exceptions to rules, whereas such effects are not easily accounted for by models using other combinations of representation and learning method.     

U-shaped learning and frequency effects in a multi-layered perceptron: implications for child language acquisition

A three-layer back-propagation network is used to implement a pattern association task in which four types of mapping are learned. These mappings, which are considered analogous to those which characterize the relationship between the stem and past tense forms of English verbs, include arbitrary mappings, identity mappings, vowel changes, and additions of a suffix. The degree of correspondence between parallel distributed processing (PDP) models which learn mappings of this sort (e.g., Rumelhart & McClelland, 1986, 1987) and children's acquisition of inflectional morphology has recently been at issue in discussions of the applicability of PDP models to the study of human cognition and language (Pinker & Mehler, 1989; Bever, in press). In this paper, we explore the capacity of a network to learn these types of mappings, focusing on three major issues. First, we compare the performance of a single-layered perceptron similar to the one used by Rumelhart and McClelland with a multi-layered perceptron. The results suggest that it is unlikely that a single-layered perceptron is capable of finding an adequate solution to the problem of mapping stems and past tense forms in input configurations that are sufficiently analogous to English. Second, we explore the input conditions which determine learning in these networks. Several factors that characterize linguistic input are investigated: (a) the nature of the mapping performed by the network (arbitrary, suffixation, identity, and vowel change); (b) the competition effects that arise when the task demands simultaneous learning of distinct mapping types; (c) the role of the type and token frequency of verb stems; and (d) the influence of phonological subregularities in the irregular verbs. Each of these factors is shown to have selective consequences on both successful and erroneous performance in the network. Third, we outline several types of systems which could result in U-shaped acquisition, and discuss the ways in which learning in multi-layered networks can be seen to capture several characteristics of U-shaped learning in children. In general, these models provide information about the role of input in determining the kinds of errors that a network will produce, including the conditions under which rule-like behavior and U-shaped learning will and will not emerge. The results from all simulations are discussed in light of behavioral data on children's acquisition of the past tense and the validity of drawing conclusions about the acquisition of language from models of this sort.     

Modeling parallelization and flexibility improvements in skill acquisition: from dual tasks to complex dynamic skills

Emerging parallel processing and increased flexibility during the acquisition of cognitive skills form a combination that is hard to reconcile with rule-based models that often produce brittle behavior. Rule-based models can exhibit these properties by adhering to 2 principles: that the model gradually learns task-specific rules from instructions and experience, and that bottom-up processing is used whenever possible. In a model of learning perfect time-sharing in dual tasks (Schumacher et al., 2001), speedup learning and bottom-up activation of instructions can explain parallel behavior. In a model of a complex dynamic task (Carnegie Mellon University Aegis Simulation Program [CMU-ASP], Anderson et al., 2004), parallel behavior is explained by the transition from serially organized instructions to rules that are activated by both top-down (goal-driven) and bottom-up (perceptually driven) factors. Parallelism lets the model opportunistically reorder instructions, leading to the gradual emergence of new task strategies.     

An interactive Hebbian account of lexically guided tuning of speech perception

We describe an account of lexically guided tuning of speech perception based on interactive processing and Hebbian learning. Interactive feedback provides lexical information to prelexical levels, and Hebbian learning uses that information to retune the mapping from auditory input to prelexical representations of speech. Simulations of an extension of the TRACE model of speech perception are presented that demonstrate the efficacy of this mechanism. Further simulations show that acoustic similarity can account for the patterns of speaker generalization. This account addresses the role of lexical information in guiding both perception and learning with a single set of principles of information propagation.     

Effects of feedback reliability on feedback-related brain activity: A feedback valuation account

Adaptive decision making relies on learning from feedback. Because feedback sometimes can be misleading, optimal learning requires that knowledge about the feedback’s reliability be utilized to adjust feedback processing. Although previous research has shown that feedback reliability indeed influences feedback processing, the underlying mechanisms through which this is accomplished remain unclear. Here we propose that feedback processing is adjusted by the adaptive, top-down valuation of feedback. We assume that unreliable feedback is devalued relative to reliable feedback, thus reducing the reward prediction errors that underlie feedback-related brain activity and learning. A crucial prediction of this account is that the effects of feedback reliability are susceptible to contrast effects. That is, the effects of feedback reliability should be enhanced when both reliable and unreliable feedback are experienced within the same context, as compared to when only one level of feedback reliability is experienced. To evaluate this prediction, we measured the event-related potentials elicited by feedback in two experiments in which feedback reliability was varied either within or between blocks. We found that the fronto-central valence effect, a correlate of reward prediction errors during reinforcement learning, was reduced for unreliable feedback. But this result was obtained only when feedback reliability was varied within blocks, thus indicating a contrast effect. This suggests that the adaptive valuation of feedback is one mechanism underlying the effects of feedback reliability on feedback processing.     

Combining exemplar-based category representations and connectionist learning rules.

Adaptive network and exemplar-similarity models were compared on their ability to predict category learning and transfer data. An exemplar-based network (Kruschke, 1990a, 1990b, 1992) that combines key aspects of both modeling approaches was also tested. The exemplar-based network incorporates an exemplar-based category representation in which exemplars become associated to categories through the same error-driven, interactive learning rules that are assumed in standard adaptive networks. Experiment 1, which partially replicated and extended the probabilistic classification learning paradigm of Gluck and Bower (1988a), demonstrated the importance of an error-driven learning rule. Experiment 2, which extended the classification learning paradigm of Medin and Schaffer (1978) that discriminated between exemplar and prototype models, demonstrated the importance of an exemplar-based category representation. Only the exemplar-based network accounted for all the major qualitative phenomena; it also achieved good quantitative predictions of the learning and transfer data in both experiments.     

汉语句子格角色分配的一种神经网络方法

提出了一个基于分布式表征的计算模型,通过并行分布加工方式完成六类汉语句子的格角色分配任务。模型是一个四层的前传网络,包括输入层（词的分布式表征层）,两个隐层,输出层（格角色层）;其中第一隐层的一部分反馈到输入层。模型采用误差反传算法,通过提供学习样本和目标输出,不断调整三个权值矩阵,使得网络稳定时能得到正确的结果。经过训练后的网络具有一定的稳定性和鲁棒性。还对这种方法与传统的符号处理方法作了比较和分析。     

Modulating semantic feedback in visual word recognition

According to the interactive activation framework proposed by McClelland and Rumelhart (1981), activation spreads both forward and backward between some levels of representation during visual word recognition. An important boundary condition, however, is that the spread of activation from lower to higher levels can be prevented (e.g., explicit letter processing during prime processing eliminates the well-documented semantic priming effect). Can the spread of activation from higher to lower levels also be prevented? This question was addressed with a choice task procedure in which subjects read a prime word and then responded to a target, performing either lexical decision or letter search depending on the color of the target. A semantic context effect was observed in lexical decision, providing evidence of semantic-level activation. In contrast, there was no semantic context effect in the letter search task, despite evidence of lexical involvement: Words were searched faster than nonwords. Further evidence of lexical involvement in the letter search task appeared in Experiment 2 in the form of greater identity priming for words than for nonwords. The results of these experiments are consistent with the conclusion that feedback from the semantic level to the lexical level can be blocked. Hence, between-level activation blocks can be instantiated in both bottom-up and top-down directions.     

Determinants of competitive performance expectancies of young male athletes1

Although performance expectancies have been shown to be important mediators of achievement behavior, few specific determinants of personal expectancies have been identified, the ecological validity of previous results is limited, and factors influencing group performance expectancies have not been determined. Therefore, the purpose of this field study involving 11- and 12-year-old male soccer players was twofold. First, we examined specific intrapersonal and situational factors influencing players' pregame personal and team performance expectancies. Second, we investigated the effects of winning and losing a competitive soccer game on players' postgame team expectancies involving a hypothetical rematch with the identical opponent. The pregame findings revealed that (a) the intrapersonal factors of ability and self-esteem were related to personal performance expectancies while competitive trait anxiety was not, and (b) the situational factors of past win-loss record and a prior game win or loss against the same opponent influenced team performance expectancies. The postgame findings revealed that future expectancies were affected by the interactive effects of game win-loss and self-esteem.     

The link between brain learning, attention, and consciousness

The processes whereby our brains continue to learn about a changing world in a stable fashion throughout life are proposed to lead to conscious experiences. These processes include the learning of top-down expectations, the matching of these expectations against bottom-up data, the focusing of attention upon the expected clusters of information, and the development of resonant states between bottom-up and top-down processes as they reach an attentive consensus between what is expected and what is there in the outside world. It is suggested that all conscious states in the brain are resonant states and that these resonant states trigger learning of sensory and cognitive representations. The models which summarize these concepts are therefore called Adaptive Resonance Theory, or ART, models. Psychophysical and neurobiological data in support of ART are presented from early vision, visual object recognition, auditory streaming, variable-rate speech perception, somatosensory perception, and cognitive-emotional interactions, among others. It is noted that ART mechanisms seem to be operative at all levels of the visual system, and it is proposed how these mechanisms are realized by known laminar circuits of visual cortex. It is predicted that the same circuit realization of ART mechanisms will be found in the laminar circuits of all sensory and cognitive neocortex. Concepts and data are summarized concerning how some visual percepts may be visibly, or modally, perceived, whereas amodal percepts may be consciously recognized even though they are perceptually invisible. It is also suggested that sensory and cognitive processing in the What processing stream of the brain obey top-down matching and learning laws that are often complementary to those used for spatial and motor processing in the brain's Where processing stream. This enables our sensory and cognitive representations to maintain their stability as we learn more about the world, while allowing spatial and motor representations to forget learned maps and gains that are no longer appropriate as our bodies develop and grow from infanthood to adulthood. Procedural memories are proposed to be unconscious because the inhibitory matching process that supports these spatial and motor processes cannot lead to resonance.     

Temporal Integration of Information in Orthographic Priming

We present a theoretical framework for studying the integration of information over time in experimental paradigms involving the successive presentation of printed strings of letters with varying orthographic overlap (orthographic priming). This framework is expressed as a set of six processing principles, five of which are derived directly from the interactive activation model of orthographic processing (McClelland & Rumelhart, 1981), and an additional sixth principle relative to a hypothesized inhibitory resetmechanism. The validity of these principles are evaluated with respect to data obtained from different variants of the orthographic priming paradigm, and related paradigms.     

Implementations are not conceptualizations: revising the verb learning model.

In a recent issue of this journal, Pinker and Prince (1988) and Lachter and Bever (1988) presented detailed critiques of Rumelhart and McClelland's (1986) connectionist model of the child's learning of the phonological form of the English past tense. In order to address these criticisms, a new connectionist model was constructed using the back-propagation algorithm, a larger input corpus, a fuller paradigm, and a new phonological representation. This new implementation successfully addressed the criticisms of the phonological representation used by Rumelhart and McClelland. It did a much better job of learning the past tense using a fuller input set with realistic frequencies of occurrence. Ancillary simulations using the same network were able to deal with the homonymy problem and the generation of forms like "ated" from "ate". The one feature not provided by the new model was a way of modeling early correct production of irregular forms. The success of the new model can be used to help clarify the extent to which the published critiques apply to a particular connectionist implementation as opposed to fundamental principles underlying the broader connectionist conceptualization.     

句法加工中的预测现象及实验证据

句法加工是读者通过整合输入的言语信息来理解句子的过程。句法加工的存在,使得读者能够提前预测文本的信息,提高阅读效率。由于阅读离不开信息整合,研究者提出两类不同的句法理论:句法预测的模块局域性理论和句法预测的互动分析性观点。这两类理论的争论点在于句法加工是单独模块化的自上而下加工,还是自下而上-自上而下的互动分析加工过程。其区别表现为句法预测的作用——前者认为句法预测对句子加工有抑制作用,后者认为句法预测对句子加工起促进作用。来自眼动、ERP等研究的数据佐证了句法预测的存在。未来研究应围绕其差异进行深入探讨,以揭示句法加工的实质。     

Coping-related expectancies and dispositions as prospective predictors of coping responses and symptoms

We used Rotter's (1954, 1982) social learning theory and Kirsch's (1985, 1999) response expectancy extension thereof to clarify distinctions between coping-related expectancies (beliefs about the outcomes of coping efforts) and coping dispositions (tendencies to use particular coping responses), specifically focusing on the role of generalized expectancies for negative mood regulation (NMR) as a predictor of individual differences in coping and well-being. Two studies using structural equation modeling provided support for direct and indirect associations between NMR expectancies and symptoms of depression. In Study 1 NMR expectancies predicted situational avoidance coping responses and symptoms of depression and anxiety, independent of dispositional avoidance coping tendencies. In Study 2, NMR expectancies were associated with depressive symptoms, concurrently and prospectively, independent of dispositional optimism and pessimism. Both studies indicated that NMR expectancies are more strongly associated with depressive symptoms than with symptoms of anxiety and physical illness. Results underscore the importance of distinguishing between expectancies and other personality variables related to coping.     

Learning Continuous Probability Distributions with Symmetric Diffusion Networks

In this article we present symmetric diffusion networks, a family of networks that instantiate the principles of continuous, stochastic, adaptive and interactive propagation of information. Using methods of Markovion diffusion theory, we formalize the activation dynamics of these networks and then show that they can be trained to reproduce entire multivariate probability distributions on their outputs using the contrastive Hebbion learning rule (CHL). We show that CHL performs gradient descent on an error function that captures differences between desired and obtained continuous multivariate probability distributions. This allows the learning algorithm to go beyond expected values of output units and to approximate complete probability distributions on continuous multivariate activation spaces. We argue that learning continuous distributions is an important task underlying a variety of real-life situations that were beyond the scope of previous connectionist networks. Deterministic networks, like back propagation, cannot learn this task because they are limited to learning average values of independent output units. Previous stochastic connectionist networks could learn probability distributions but they were limited to discrete variables. Simulations show that symmetric diffusion networks can be trained with the CHL rule to approximate discrete and continuous probability distributions of various types.     

Recent models and findings in visual backward masking: a comparison, review, and update

Visual backward masking not only is an empirically rich and theoretically interesting phenomenon but also has found increasing application as a powerful methodological tool in studies of visual information processing and as a useful instrument for investigating visual function in a variety of specific subject populations. Since the dual-channel, sustained-transient approach to visual masking was introduced about two decades ago, several new models of backward masking and metacontrast have been proposed as alternative approaches to visual masking. In this article, we outline, review, and evaluate three such approaches: an extension of the dual-channel approach as realized in the neural network model of retino-cortical dynamics (Ogmen, 1993), the perceptual retouch theory (Bachmann, 1984, 1994), and the boundary contour system (Francis, 1997; Grossberg & Mingolla, 1985b). Recent psychophysical and electrophysiological findings relevant to backward masking are reviewed and, whenever possible, are related to the aforementioned models. Besides noting the positive aspects of these models, we also list their problems and suggest changes that may improve them and experiments that can empirically test them.