Publications

Abstract

Neuron models with explicit dendritic dynamics have shed light on mechanisms for coincidence detection, pathway selection and temporal filtering. However, it is still unclear which morphological and physiological features are required to capture these phenomena. In this work, we introduce the Tripod neuron model and propose a minimal structural reduction of the dendritic tree that is able to reproduce these computations. The Tripod is a three-compartment model consisting of two segregated passive dendrites and a somatic compartment modelled as an adaptive, exponential integrate-and-fire neuron. It incorporates dendritic geometry, membrane physiology and receptor dynamics as measured in human pyramidal cells. We characterize the response of the Tripod to glutamatergic and GABAergic inputs and identify parameters that support supra-linear integration, coincidence-detection and pathway-specific gating through shunting inhibition. Following NMDA spikes, the Tripod neuron generates plateau potentials whose duration depends on the dendritic length and the strength of synaptic input. When fitted with distal compartments, the Tripod encodes previous activity into a dendritic depolarized state. This dendritic memory allows the neuron to perform temporal binding, and we show that it solves transition and sequence detection tasks on which a single-compartment model fails. Thus, the Tripod can account for dendritic computations previously explained only with more detailed neuron models or neural networks. Due to its simplicity, the Tripod neuron can be used efficiently in simulations of larger cortical circuits.

Abstract

Language processing involves the ability to store and integrate pieces of
information in working memory over short periods of time. According to
the dominant view, information is maintained through sustained, elevated
neural activity. Other work has argued that short-term synaptic facilitation
can serve as a substrate of memory. Here, we propose an account where
memory is supported by intrinsic plasticity that downregulates neuronal
firing rates. Single neuron responses are dependent on experience and we
show through simulations that these adaptive changes in excitability pro-
vide memory on timescales ranging from milliseconds to seconds. On this
account, spiking activity writes information into coupled dynamic variables
that control adaptation and move at slower timescales than the membrane
potential. From these variables, information is continuously read back into
the active membrane state for processing. This neuronal memory mech-
anism does not rely on persistent activity, excitatory feedback, or synap-
tic plasticity for storage. Instead, information is maintained in adaptive
conductances that reduce firing rates and can be accessed directly with-
out cued retrieval. Memory span is systematically related to both the time
constant of adaptation and baseline levels of neuronal excitability. Inter-
ference effects within memory arise when adaptation is long-lasting. We
demonstrate that this mechanism is sensitive to context and serial order
which makes it suitable for temporal integration in sequence processing
within the language domain. We also show that it enables the binding of
linguistic features over time within dynamic memory registers. This work
provides a step towards a computational neurobiology of language.

Abstract

Event-related potentials (ERPs) provide a window into how the brain is processing language. Here, we propose a theory that argues that ERPs such as the N400 and P600 arise as side effects of an error-based learning mechanism that explains linguistic adaptation and language learning. We instantiated this theory in a connectionist model that can simulate data from three studies on the N400 (amplitude modulation by expectancy, contextual constraint, and sentence position), five studies on the P600 (agreement, tense, word category, subcategorization and garden-path sentences), and a study on the semantic P600 in role reversal anomalies. Since ERPs are learning signals, this account explains adaptation of ERP amplitude to within-experiment frequency manipulations and the way ERP effects are shaped by word predictability in earlier sentences. Moreover, it predicts that ERPs can change over language development. The model provides an account of the sensitivity of ERPs to expectation mismatch, the relative timing of the N400 and P600, the semantic nature of the N400, the syntactic nature of the P600, and the fact that ERPs can change with experience. This approach suggests that comprehension ERPs are related to sentence production and language acquisition mechanisms

Abstract

Biologically inspired spiking networks are an important tool to study the nature of computation and cognition in neural systems. In this work, we investigate the representational capacity of spiking networks engaged in an identity mapping task. We compare two schemes for encoding symbolic input, one in which input is injected as a direct current and one where input is delivered as a spatio-temporal spike pattern. We test the ability of networks to discriminate their input as a function of the number of distinct input symbols. We also compare performance using either membrane potentials or filtered spike trains as state variable. Furthermore, we investigate how the circuit behavior depends on the balance between excitation and inhibition, and the degree of synchrony and regularity in its internal dynamics. Finally, we compare different linear methods of decoding population activity onto desired target labels. Overall, our results suggest that even this simple mapping task is strongly influenced by design choices on input encoding, state-variables, circuit characteristics and decoding methods, and these factors can interact in complex ways. This work highlights the importance of constraining computational network models of behavior by available neurobiological evidence.

Abstract

Nativist theories have argued that language involves syntactic principles which are unlearnable from the input children receive. A paradigm case of these innate principles is the structure dependence of auxiliary inversion in complex polar questions (Chomsky, 1968, 1975, 1980). Computational approaches have focused on the properties of the input in explaining how children acquire these questions. In contrast, we argue that messages are structured in a way that supports structure dependence in syntax. We demonstrate this approach within a connectionist model of sentence production (Chang, 2009) which learned to generate a range of complex polar questions from a structured message without positive exemplars in the input. The model also generated different types of error in development that were similar in magnitude to those in children (e.g., auxiliary doubling, Ambridge, Rowland, & Pine, 2008; Crain & Nakayama, 1987). Through model comparisons we trace how meaning constraints and linguistic experience interact during the acquisition of auxiliary inversion. Our results suggest that auxiliary inversion rules in English can be acquired without innate syntactic principles, as long as it is assumed that speakers who ask complex questions express messages that are structured into multiple propositions

Abstract

Prior language input is not lost but integrated with the current input. This principle is demonstrated by “reservoir computing”: Untrained recurrent neural networks project input sequences onto a random point in high-dimensional state space. Earlier inputs can be retrieved from this projection, albeit less reliably so as more input is received. The bottleneck is therefore not “Now-or-Never” but “Sooner-is-Better.

Abstract

Rey et al. (2012) present data from a study with baboons that they interpret in support of the idea that center-embedded structures in human language have their origin in low level memory mechanisms and associative learning. Critically, the authors claim that the baboons showed a behavioral preference that is consistent with center-embedded sequences over other types of sequences. We argue that the baboons’ response patterns suggest that two mechanisms are involved: first, they can be trained to associate a particular response with a particular stimulus, and, second, when faced with two conditioned stimuli in a row, they respond to the most recent one first, copying behavior they had been rewarded for during training. Although Rey et al. (2012) ‘experiment shows that the baboons’ behavior is driven by low level mechanisms, it is not clear how the animal behavior reported, bears on the phenomenon of Center Embedded structures in human syntax. Hence, (1) natural language syntax may indeed have been shaped by low level mechanisms, and (2) the baboons’ behavior is driven by low level stimulus response learning, as Rey et al. propose. But is the second evidence for the first? We will discuss in what ways this study can and cannot give evidential value for explaining the origin of Center Embedded recursion in human grammar. More generally, their study provokes an interesting reflection on the use of animal studies in order to understand features of the human linguistic system.

Abstract

Lexicalized theories of syntax often assume that verb-structure regularities are mediated by lemmas, which abstract over variation in verb tense and aspect. German syntax seems to challenge this assumption, because verb position depends on tense and aspect. To examine how German speakers link these elements, a structural priming study was performed which varied syntactic structure, verb position (encoded by tense and aspect), and verb overlap. Abstract structural priming was found, both within and across verb position, but priming was larger when the verb position was the same between prime and target. Priming was boosted by verb overlap, but there was no interaction with verb position. The results can be explained by a lemma model where tense and aspect are linked to structural choices in German. Since the architecture of this lemma model is not consistent with results from English, a connectionist model was developed which could explain the cross-linguistic variation in the production system. Together, these findings support the view that language learning plays an important role in determining the nature of structural priming in different languages

Abstract

Sentence production is the process we use to create language-specific sentences that convey particular meanings. In production, there are complex interactions between meaning, words, and syntax at different points in sentences. Computational models can make these interactions explicit and connectionist learning algorithms have been useful for building such models. Connectionist models use domaingeneral mechanisms to learn internal representations and these mechanisms can also explain evidence of long-term syntactic adaptation in adult speakers. This paper will review work showing that these models can generalize words in novel ways and learn typologically-different languages like English and Japanese. It will also present modeling work which shows that connectionist learning algorithms can account for complex sentence production in children and adult production phenomena like structural priming, heavy NP shift, and conceptual/lexical accessibility.

Abstract

Relative clauses are a syntactic device to create complex sentences and they make language structurally productive. Despite a considerable number of experimental studies, it is still largely unclear how children learn relative clauses and how these are processed in the language system. Researchers at the MPI for Psycholinguistics used a computational learning model to gain novel insights into these issues. The model explains the differential development of relative clauses in English as well as cross-linguistic differences

Abstract

In traditional theories of language comprehension, syntactic and semantic processing are inextricably linked. This assumption has been challenged by the ‘Semantic Illusion Effect’ found in studies using Event Related brain Potentials. Semantically anomalous sentences did not produce the expected increase in N400 amplitude but rather one in P600 amplitude. To explain these findings, complex models have been devised in which an independent semantic processing stream can arrive at a sentence interpretation that may differ from the interpretation prescribed by the syntactic structure of the sentence. We review five such multi-stream models and argue that they do not account for the full range of relevant results because they assume that the amplitude of the N400 indexes some form of semantic integration. Based on recent evidence we argue that N400 amplitude might reflect the retrieval of lexical information from memory. On this view, the absence of an N400-effect in Semantic Illusion sentences can be explained in terms of priming. Furthermore, we suggest that semantic integration, which has previously been linked to the N400 component, might be reflected in the P600 instead. When combined, these functional interpretations result in a single-stream account of language processing that can explain all of the Semantic Illusion data.

Abstract

Linguistic adaptation is a phenomenon where language representations change in response to linguistic input. Adaptation can occur on multiple linguistic levels such as phonology (tuning of phonotactic constraints), words (repetition priming), and syntax (structural priming). The persistent nature of these adaptations suggests that they may be a form of implicit learning and connectionist models have been developed which instantiate this hypothesis. Research on implicit learning, however, has also produced evidence that explicit chunk knowledge is involved in the performance of these tasks. In this review, we examine how these interacting implicit and explicit processes may change our understanding of language learning and processing.

Abstract

Relative clause processing depends on the grammatical role of the head noun in the subordinate clause. This has traditionally been explained in terms of cognitive limitations. We suggest that structure-related processing differences arise from differences in experience with these structures. We present a connectionist model which learns to produce utterances with relative clauses from exposure to message-sentence pairs. The model shows how various factors such as frequent subsequences, structural variations, and meaning conspire to create differences in the processing of these structures. The predictions of this learning-based account have been confirmed in behavioral studies with adults. This work shows that structural regularities that govern relative clause processing can be explained within a usage-based approach to recursion.

Abstract

Language acquisition involves learning nonadjacent dependencies that can obtain between words in a sentence. Several artificial grammar learning studies have shown that the ability of adults and children to detect dependencies between A and B in frames AXB is influenced by the amount of variation in the X element. This paper presents a model of statistical learning which displays similar behavior on this task and generalizes in a human-like way. The model was also used to predict human behavior for increased distance and more variation in dependencies. We compare our model-based approach with the standard invariance account of the variability effect.

Abstract

This paper investigates whether surprisal theory can account for differential processing difficulty in the NP-/S-coordination ambiguity in Dutch. Surprisal is estimated using a Probabilistic Context-Free Grammar (PCFG), which is induced from an automatically annotated corpus. We find that our lexicalized surprisal model can account for the reading time data from a classic experiment on this ambiguity by Frazier (1987). We argue that syntactic and lexical probabilities, as specified in a PCFG, are sufficient to account for what is commonly referred to as an NP-coordination preference.

Abstract

The problem of auxiliary fronting in complex polar questions occupies a prominent position within the nature versus nurture controversy in language acquisition. We employ a model of statistical learning which uses sequential and semantic information to produce utterances from a bag of words. This linear learner is capable of generating grammatical questions without exposure to these structures in its training environment. We also demonstrate that the model performs superior to n-gram learners on this task. Implications for nativist theories of language acquisition are discussed.

Abstract

Children learn their mother tongue spontaneously and effortlessly through communicative interaction with their environment; they do not have to be taught explicitly or learn how to learn first. The ambient language to which children are exposed, however, is highly variable and arguably deficient with regard to the learning target. Nonetheless, most normally developing children learn their native language rapidly and with ease. To explain this accomplishment, many theories of acquisition posit innate constraints on learning, or even a biological endowment for language which is specific to language. Usage-based theories, on the other hand, place more emphasis on the role of experience and domain-general learning mechanisms than on innate language-specific knowledge. But languages are lexically open and combinatorial in structure, so no amount of experience covers their expressivity. Usage-based theories therefore have to explain how children can generalize the properties of their linguistic input to an adult-like grammar. In this thesis I provide an explicit computational mechanism with which usage-based theories of language can be tested and evaluated. The focus of my work lies on complex syntax and the human ability to form sentences which express more than one proposition by means of relativization. This `capacity for recursion' is a hallmark of an adult grammar and, as some have argued, the human language faculty itself. The manuscript is organized as follows. In the second chapter, I give an overview of results that characterize the properties of neural networks as mathematical objects and review previous attempts at modelling the acquisition of complex syntax with such networks. The chapter introduces the conceptual landscape in which the current work is located. In the third chapter, I argue that the construction and use of meaning is essential in child language acquisition and adult processing. Neural network models need to incorporate this dimension of human linguistic behavior. I introduce the Dual-path model of sentence production and syntactic development which is able to represent semantics and learns from exposure to sentences paired with their meaning (cf. Chang et al. 2006). I explain the architecture of this model, motivate critical assumptions behind its design, and discuss existing research using this model. The fourth chapter describes and compares several extensions of the basic architecture to accommodate the processing of multi-clause utterances. These extensions are evaluated against computational desiderata, such as good learning and generalization performance and the parsimony of input representations. A single-best solution for encoding the meaning of complex sentences with restrictive relative clauses is identified, which forms the basis for all subsequent simulations. Chapter five analyzes the learning dynamics in more detail. I first examine the model's behavior for different relative clause types. Syntactic alternations prove to be particularly difficult to learn because they complicate the meaning-to-form mapping the model has to acquire. In the second part, I probe the internal representations the model has developed during learning. It is argued that the model acquires the argument structure of the construction types in its input language and represents the hierarchical organization of distinct multi-clause utterances. The juice of this thesis is contained in chapters six to eight. In chapter six, I test the Dual-path model's generalization capacities in a variety of tasks. I show that its syntactic representations are sufficiently transparent to allow structural generalization to novel complex utterances. Semantic similarities between novel and familiar sentence types play a critical role in this task. The Dual-path model also has a capacity for generalizing familiar words to novel slots in novel constructions (strong semantic systematicity). Moreover, I identify learning conditions under which the model displays recursive productivity. It is argued that the model's behavior is consistent with human behavior in that production accuracy degrades with depth of embedding, and right-branching is learned faster than center-embedding recursion. In chapter seven, I address the issue of learning complex polar interrogatives in the absence of positive exemplars in the input. I show that the Dual-path model can acquire the syntax of these questions from simpler and similar structures which are warranted in a child's linguistic environment. The model's errors closely match children's errors, and it is suggested that children might not require an innate learning bias to acquire auxiliary fronting. Since the model does not implement a traditional kind of language-specific universal grammar, these results are relevant to the poverty of the stimulus debate. English relative clause constructions give rise to similar performance orderings in adult processing and child language acquisition. This pattern matches the typological universal called the noun phrase accessibility hierarchy. I propose an input-based explanation of this data in chapter eight. The Dual-path model displays this ordering in syntactic development when exposed to plausible input distributions. But it is possible to manipulate and completely remove the ordering by varying properties of the input from which the model learns. This indicates, I argue, that patterns of interference and facilitation among input structures can explain the hierarchy when all structures are simultaneously learned and represented over a single set of connection weights. Finally, I draw conclusions from this work, address some unanswered questions, and give a brief outlook on how this research might be continued.

Abstract

We present a neural-symbolic learning model of sentence production which displays strong semantic systematicity and recursive productivity. Using this model, we provide evidence for the data-driven learnability of complex yes/no- questions.

Abstract

We argue that van der Velde's & de Kamps's model does not solve the binding problem but merely shifts the burden of constructing appropriate neural representations of sentence structure to unexplained preprocessing of the linguistic input. As a consequence, their model is not able to explain how various neural representations can be assigned to sentences that are structurally ambiguous.