Franklin, J. A. (2006). Jazz melody generation using recurrent networks and reinforcement learning. International Journal on Artificial Intelligence Tools, 15(4), 623-650.
Recurrent (neural) networks have been deployed as models for learning musical processes, by computational scientists who study processes such as dynamic systems. Over time, more intricate music has been learned as the state of the art in recurrent networks improves. One particular recurrent network, the Long Short-Term Memory (LSTM) network shows promise for learning long songs, and generating new songs. We are experimenting with a module containing two inter-recurrent LSTM networks to cooperatively learn several human melodies, based on, the songs' harmonic structures, and on the feedback inherent in the network. We show that these networks can learn to reproduce four human melodies. We then present as input new harmonizations, so as to generate new songs. We describe the reharmonizations, and show the new melodies that result. We also present a hierarchical structure for using reinforcement learning to choose LSTM modules during the course of melody generation.
Gurney, K. (2007). Neural networks for perceptual processing: from simulation tools to theories. Philosophical Transactions of the Royal Society B-Biological Sciences, 362(1479), 339-353.
Neural networks are modelling tools that are, in principle, able to capture the input-output behaviour of arbitrary systems that may include the dynamics of animal populations or brain circuits. While a neural network model is useful if it captures phenomenologically the behaviour of the target system in this way, its utility is amplified if key mechanisms of the model can be discovered, and identified with those of the underlying system. In this review, we first describe, at a fairly high level with minimal mathematics, some of the tools used in constructing neural network models. We then go on to discuss the implications of network models for our understanding of the system they are supposed to describe, paying special attention to those models that deal with neural circuits and brain systems. We propose that neural nets are useful for brain modelling if they are viewed in a wider computational framework originally devised by Marr. Here, neural networks are viewed as an intermediate mechanistic abstraction between 'algorithm' and 'implementation', which can provide insights into biological neural representations and their putative supporting architectures.
Hartshorne, J. K., & Ullman, M. T. (2006). Why girls say 'holded' more than boys. Developmental Science, 9(1), 21-32.
Women are better than men at verbal memory tasks, such as remembering word lists. These tasks depend on declarative memory. The declarative/procedural model of language, which posits that the lexicon of stored words is part of declarative memory, while grammatical composition of complex forms depends on procedural memory, predicts a female superiority in aspects of lexical memory. Other neurocognitive models of language have not made this prediction. Here we examine the prediction in past-tense over-regularizations (e.g holded) produced by children. We expected that girls would remember irregular past-tense forms (held) better than boys, and thus would over-regularize less. To our surprise, girls over-regularized far more than boys. We investigated potential explanations for this sex difference. Analyses showed that in girls but not boys, over-regularization rates correlated with measures of the number of similar-sounding regulars (folded, molded). This sex difference in phonological neighborhood effects is taken to suggest that girls tend to produce over-regularizations in associative lexical memory, generalizing over stored neighboring regulars, while boys are more likely to depend upon rule-governed affixation (hold + -ed). The finding is consistent with the hypothesis that, likely due to their superior lexical abilities, females tend to retrieve from memory complex forms (walked) that men generally compose with the grammatical system (walk + -ed). The results suggest that sex may be an important factor in the acquisition and computation of language.
Kandel, E. R. (2001). Neuroscience - The molecular biology of memory storage: A dialogue between genes and synapses. Science, 294(5544), 1030-1038.
One of the most remarkable aspects of an animal's behavior is the ability to modify that behavior by learning, an ability that reaches its highest form in human beings. For me, learning and memory have proven to be endlessly fascinating mental processes because they address one of the fundamental features of human activity: our ability to acquire new ideas from experience and to retain these ideas over time in memory. Moreover, unlike other mental processes such as thought, language, and consciousness, learning seemed from the outset to be readily accessible to cellular and molecular analysis. I, therefore, have been curious to know: What changes in the brain when we learn? And, once something is learned, how is that information retained in the brain? I have tried to address these questions through a reductionist approach that would allow me to investigate elementary forms of learning and memory at a cellular molecular level-as specific molecular activities within identified nerve cells.
Marshall, C. R., & van der Lely, H. K. J. (2006). A challenge to current models of past tense inflection: The impact of phonotactics. Cognition, 100(2), 302-320.
Is past tense production better modelled by a Single Mechanism or a Words and Rules model? We present data concerning a phenomenon that has not been considered by either model-regular past tense verbs with contrasting phonotactics. One set of verbs contains clusters at the inflected verb end that also occur in monomorphemic words ('monomorphemically legal clusters', MLC) whereas the other has clusters that can only occur in inflected forms ('monomorphemically illegal clusters', MIC). We argue that if children apply a morphological rule, phonotactics will not affect performance. Conversely, if children store past tense forms, they will perform better on verbs with MLCs because these clusters are more frequent. We investigated three populations-typically developing children, Grammatical-SLI (G-SLI) and Williams Syndrome (WS)-using past tense elicitation tasks. In Experiment 1 we reanalyse data from van der Lely and Ullman [van der Lely, H. K. J. & Ullman, M. (2001). Past tense morphology in specifically language impaired and normally developing children. Language and Cognitive Processes, 16: 177-217] and show that G-SLI children perform better on MLC verbs, whereas for typically developing children phonotactics do not affect performance. In Experiment 2 we replicate these findings in new groups of G-SLI and typically developing children. In Experiment 3 we reanalyse data from Thomas et al. [Thomas, M. S. C., Grant, J., Barham, Z., Gsodl, M., Laing, E., Lakusta, L., Tyler, L.K., Grice, S., Paterson, S. & Karmiloff-Smith, A. (2001) Past tense formation in Williams Syndrome. Language and Cognitive Processes, 16: 143-176] and show that phonotactics do not affect performance in individuals with WS. We argue that the results elucidate the underlying nature of morphology in these populations, and are better accommodated within a Words and Rules model of past tense acquisition. (c) 2005 Elsevier B.V. All rights reserved.
[not in handout, see intranet]
Marslen-Wilson, W., & Tyler, L. (2007). Morphology, language and the brain: the decompositional substrate for language comprehension. Philosophical Transactions of the Royal Society B: Biological Sciences, 362(1481), 823-836.
This emphasis on the decomposition of regular inflected forms clearly allies the account presented here with the ‘Words and Rules’ (e.g. Clahsen 1999; Pinker 1999) and the procedural/declarative (e.g. Ullman 2004) approaches, in distinction to nondecompositional, usually connectionist approaches, which deny the existence of separable stem and inflectional morphemes and argue instead that inflected forms are processed and represented as overlapping whole forms sharing certain semantic and phonological similarities (e.g. Rumelhart & McClelland 1986; McClelland & Patterson 2002).
Monaghan, P., & Shillcock, R. (2007). Levels of description in consonant/vowel processing: Reply to Knobel and Caramazza. Brain and Language, 100(1), 101-108.
Is it necessary to posit separate, explicit distinctions between representations in order to account for dissociations between consonant and vowel processing? We argue that a cognitive model of speech production based on cumulative lower-level properties is not only sufficient but more parsimonious in accounting for aphasic and dysgraphic patient data. We re-examine a computational model of consonant and vowel processing based oil phonological feature representations of phonemes, and show that models based oil similar principles are sufficient to account for quantitative and qualitative aspects of the patient data. We argue that the facts that aphasic patients (i) are more likely to have impairment to consonant than vowel representations, (ii) demonstrate varying degrees of impairment to both consonants and vowels in ail inverse relationship. and (iii) never indicate complete impairment to only vowels or consonants, are better accounted for in a model that assumes a continuum of representations of consonants and vowels than a model that explicitly encodes the consonant/vowel distinction. (C) 2006 Elsevier Inc. All rights reserved.
[not in handout, see intranet]
Newman et al. (2007). An ERP study of regular and irregular English past tense inflection. Neuroimage, 34(1), 435-445.
Compositionality is a critical and universal characteristic of human language. It is found at numerous levels, including the combination of morphemes into words and of words into phrases and sentences. These compositional patterns can generally be characterized by rules. For example, the past tense of most English verbs ("regulars") is formed by adding an -ed suffix. However, many complex linguistic forms have rather idiosyncratic mappings. For example, "irregular" English verbs have past tense forms that cannot be derived from their stems in a consistent manner. Whether regular and irregular forms depend on fundamentally distinct neurocognitive processes (rule-governed combination vs. lexical memorization), or whether a single processing system is sufficient to explain the phenomena, has engendered considerable investigation and debate. We recorded event-related potentials while participants read English sentences that were either correct or had violations of regular past tense inflection, irregular past tense inflection, syntactic phrase structure, or lexical semantics. Violations of regular past tense and phrase structure, but not of irregular past tense or lexical semantics, elicited left-lateralized anterior negativities (LANs). These seem to reflect neurocognitive substrates that underlie compositional processes across linguistic domains, including morphology and syntax. Regular, irregular, and phrase structure violations all elicited later positivities that were maximal over midline parietal sites (P600s), and seem to index aspects of controlled syntactic processing of both phrase structure and morphosyntax. The results suggest distinct neurocognitive substrates for processing regular and irregular past tense forms: regulars depending on compositional processing, and irregulars stored in lexical memory. (c) 2006 Elsevier Inc. All rights reserved.
[not in handout, see intranet]
Nicoladis, E., Palmer, A., & Marentette, P. (2007). The role of type and token frequency in using past tense morphemes correctly. Developmental Science, 10(2), 237-254.
Type and token frequency have been thought to be important in the acquisition of past tense morphology, particularly in differentiating regular and irregular forms. In this study we tested the role of frequency in two ways: (1) in bilingual children, who typically use and hear either language less often than monolingual children and (2) cross-linguistically: French and English have different patterns of frequency of regular/irregular verbs. Ten French-English bilingual children, 10 French monolingual and 10 English monolingual children between 4 and 6 years watched a cartoon and re-told the story. The results demonstrated that the bilingual children were less accurate than the monolingual children. Their accuracy in both French and English regular and irregular verbs corresponded to frequency in the input language. These results are consistent with the hypothesis that children learn past tense morphemes by analogy with other words in their vocabularies. We propose a developmental sequence based on conservative generalization across a growing set of verbs. [“The proposed developmental trajectory we have laid out assumes only one process: conservative generalization based on phonological similarity (as in Bybee, 1995; Plunkett & Marchman, 1993).”]
[not in handout, see intranet]
Penke, M., & Westermann, G. (2006). Broca's area and inflectional morphology: Evidence from Broca's aphasia and computer modeling. Cortex, 42(4), 563-576.
In a series of articles Ullman (2001, 2004; Ullman et al., 1997) has proposed that regular inflection is critically subserved by Broca's area. This suggestion is motivated by the finding that English speaking Broca's aphasics show selective deficits with regular inflection. Here we argue that this proposal does not hold cross-linguistically but is based on a confound between inflectional suffix and regularity that is specific to the English language. We present data from two experimental studies of participle inflection with 13 German and 12 Dutch Broca's aphasics. None of these aphasic speakers are selectively impaired for regular inflection but instead most of them show selective deficits with irregular inflection. These data suggest that a selective regular deficit is not a characteristic of Broca's aphasia across languages, and that Broca's area is not critically involved in regular inflection. To investigate the nature and localization of the processes underlying inflection we present a connectionist neural network model that accounts for the deficits of the German aphasic speakers. The model implements the view that the inflection of all verb types is based on a single mechanism with multiple representations that emerge from experience-dependent brain development. We show that global damage to this model results in a selective deficit for irregular inflection that is comparable to that of the German aphasic speakers. This finding suggests that a selective impairment of irregular participles as observed by German and Dutch aphasic speakers does not presuppose two distinctly localized mechanisms or processes that can be selectively affected by brain damage.
Plunkett, K., & Bandelow, S. (2006). Stochastic approaches to understanding dissociations in inflectional morphology. Brain and Language, 98(2), 194-209.
Computer modelling research has undermined the view that double dissociations in behaviour are sufficient to infer separability in the cognitive mechanisms underlying those behaviours. However, all these models employ multi-modal representational schemes, where functional specialisation of processing emerges from the training process. Targeted lesioning of different regions of functional specialisation leads to varied but predictable deficits in model performance. We argue that multi-modal representational schemes are not a necessary condition for the observation of double dissociations in an information processing system that shares resources across multiple tasks. Using a uni-modal representational system, we demonstrate that double dissociations may also result from stochastic processes. Lesioning experiments on a single-route, uni-modal connectionist model of regular and irregular noun and verb morphology confirm and extend earlier work demonstrating that selective impairment across tasks can result from damage to a distributed information processing system. A systematic investigation of the degree to which performance deteriorates across different inflectional classes reveals that simple and double dissociations can occur in this single-route, uni-modal model. An important prediction of the model is that double dissociations between regular and irregular inflection, resulting from stochastic processes should be extremely rare. However, they are particularly likely to occur when the researcher uses test batteries consisting of a small number of items. Given that cognitive neuropsychologists rarely provide details about the distribution of performance in a disordered population, it is concluded that a stochastic interpretation of double dissociations may have wider applicability than is normally supposed. (c) 2006 Elsevier Inc. All rights reserved.
Westermann, G., Sirois, S., Shultz, T. R., & Mareschal, D. (2006). Modeling developmental cognitive neuroscience. Trends in Cognitive Sciences, 10(5), 227-232.
In the past few years connectionist models have greatly contributed to formulating theories of cognitive development. Some of these models follow the approach of developmental cognitive neuroscience in exploring interactions between brain development and cognitive development by integrating structural change into learning. We describe two classes of these models. The first focuses on experience-dependent structural elaboration within a brain region by adding or deleting units and connections during learning. The second models the gradual integration of different brain areas based on combinations of experience-dependent and maturational factors. These models provide new theories of the mechanisms of cognitive change in various domains and they offer an integrated framework to study normal and abnormal development, and normal and impaired adult processing.
Westermann, G., Mareschal, D., Johnson, M. H., Sirois, S., Spratling, M. W., & Thomas, M. S. C. (2007). Neuroconstructivism. Developmental Science, 10(1), 75-83.
Neuroconstructivism is a theoretical framework focusing on the construction of representations in the developing brain. Cognitive development is explained as emerging from the experience-dependent development of neural structures supporting mental representations. Neural development occurs in the context of multiple interacting constraints acting on different levels, from the individual cell to the external environment of the developing child. Cognitive development can thus be understood as a trajectory originating from the constraints on the underlying neural structures. This perspective offers an integrated view of normal and abnormal development as well as of development and adult processing, and it stands apart from traditional cognitive approaches in taking seriously the constraints on cognition inherent to the substrate that delivers it.
White, R. L., & Snyder, L. H. (2007). Spatial constancy and the brain: insights from neural networks. Philosophical Transactions of the Royal Society B-Biological Sciences, 362(1479), 375-382.
To form an accurate internal representation of visual space, the brain must accurately account for movements of the eyes, head or body. Updating of internal representations in response to these movements is especially important when remembering spatial information, such as the location of an object, since the brain must rely on non-visual extra-retinal signals to compensate for self-generated movements. We investigated the computations underlying spatial updating by constructing a recurrent neural network model to store and update a spatial location based on a gaze shift signal, and to do so flexibly based on a contextual cue. We observed a striking similarity between the patterns of behaviour produced by the model and monkeys trained to perform the same task, as well as between the hidden units of the model and neurons in the lateral intraparietal area ( LIP). In this report, we describe the similarities between the model and single unit physiology to illustrate the usefulness of neural networks as a tool for understanding specific computations performed by the brain.