Implicit Learning and Recursion

Rohrmeier, Martin; Dienes, Zoltán; Guo, Xin-Heng; Fu, Qiuping

doi:10.1007/978-1-4614-9414-0_6

Cited by 17 publications

(15 citation statements)

References 67 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We demonstrate in this study that recurrent neural network models are closer to humans than feedforward ones, irrespective of the grammars' level in the Chomsky's hierarchy. This result endorses recurrent models as the more plausible cognitive architecture underlying language, pointing to an essential role of recursion in grammar learning (Rohrmeier, Dienes, Guo, & Fu, 2014). However, it could be fetching to generalize this conclusion to natural language acquisition (Corballis, 2014;Jackendoff, 2011;Fitch & Friederici, 2012;Vergauwen, 2014): the Chomsky hierarchy, despite being an excellent reference, does not embrace all natural grammars (Jäger & Rogers, 2012), and does not fully reflect human cognitive complexity (Öttl et al, 2015).…”

Section: Resultsmentioning

confidence: 96%

Which Neural Network Architecture matches Human Behavior in Artificial Grammar Learning?

Alamia

Gauducheau

Paisios

et al. 2019

2019 Conference on Cognitive Computational Neuroscience

View full text Add to dashboard Cite

In recent years artificial neural networks achieved performance close to or better than humans in several domains: tasks that were previously human prerogatives, such as language processing, have witnessed remarkable improvements in state of the art models. One advantage of this technological boost is to facilitate comparison between different neural networks and human performance, in order to deepen our understanding of human cognition. Here, we investigate which neural network architecture (feed-forward vs. recurrent) matches human behavior in artificial grammar learning, a crucial aspect of language acquisition. Prior experimental studies proved that artificial grammars can be learnt by human subjects after little exposure and often without explicit knowledge of the underlying rules. We tested four grammars with different complexity levels both in humans and in feedforward and recurrent networks. Our results show that both architectures can "learn" (via error back-propagation) the grammars after the same number of training sequences as humans do, but recurrent networks perform closer to humans than feedforward ones, irrespective of the grammar complexity level. Moreover, similar to visual processing, in which feedforward and recurrent architectures have been related to unconscious and conscious processes, our results suggest that explicit learning is best modeled by recurrent architectures, whereas feedforward networks better capture the dynamics involved in implicit learning.

show abstract

Section: Resultsmentioning

confidence: 96%

Which Neural Network Architecture matches Human Behavior in Artificial Grammar Learning?

Alamia

Gauducheau

Paisios

et al. 2019

2019 Conference on Cognitive Computational Neuroscience

View full text Add to dashboard Cite

show abstract

“…The questions are of course related in that certain structures require learning devices of a certain level of complexity; and a given learning device defines a class of learnable structures. One prominent way of classifying structure complexity that may be relevant to implicit learning is whether the structure could be processed by a device of finite state or rather more than finite state complexity (e.g., Poletiek, 2011 ; Rohrmeier et al, 2014 ).…”

Section: Introductionmentioning

confidence: 99%

Fluency Expresses Implicit Knowledge of Tonal Symmetry

Ling

Qiao

et al. 2016

Front. Psychol.

View full text Add to dashboard Cite

The purposes of the present study were twofold. First, we sought to establish whether tonal symmetry produces processing fluency. Second, we sought to explore whether symmetry and chunk strength express themselves differently in fluency, as an indication of different mechanisms being involved for sub- and supra-finite state processing. Across two experiments, participants were asked to listen to and memorize artificial poetry showing a mirror symmetry (an inversion, i.e., a type of cross serial dependency); after this training phase, people completed a four-choice RT task in which they were presented with new artificial poetry. Participants were required to identify the stimulus displayed. We found that symmetry sped up responding to the second half of strings, indicating a fluency effect. Furthermore, there was a dissociation between fluency effects arising from symmetry vs. chunk strength, with stronger fluency effects for symmetry rather than chunks in the second half of strings. Taken together, we conjecture a divide between finite state and supra-finite state mechanisms in learning grammatical sequences.

show abstract

“…E, that need to be encoded). Therefore, if there are multiple (finite), potentially nested non-local dependencies the number of required states grows exponentially (see also the comparable argument regarding the implicit acquisition of such structures in references [99,114]). …”

Section: Moving Towards Different Types Of Modelsmentioning

confidence: 99%

“…One aspect stems from the fact that the CH is by its definition fundamentally tied to rewrite rules and the structures that different types of rewrite rules constrained by different restrictions may express. One well-known issue-and an aspect that the notion of mild context-sensitivity addresses-concerns the fact that repetition, repetition under a modification (such as musical transposition) and cross-serial dependencies constitute types of structures that require quite complex rewrite rules (see also the example of context-sensitive rewrite rules expressing cross-serial dependencies in reference [114]). In contrast, such phenomena are frequent forms of formbuilding in music [25] and animal song.…”

Section: Limitations Of the Chomsky Hierarchymentioning

confidence: 99%

“…In particular, a number of linguistic formalisms (e.g. tree-adjoining languages, combinatorial categorial grammar [110,111]) have been proposed that have been collectively referred to as mildly context-sensitive [112]. These grammars define classes of languages that are subsets of context-sensitive languages and supersets of the context-free languages.…”

Section: Limitations Of the Chomsky Hierarchymentioning

confidence: 99%

See 1 more Smart Citation