Structured Sequence Processing and Combinatorial Binding: Neurobiologically Royalsocietypublishing.Org/Journal/Rstb and Computationally Informed Hypotheses

Structured Sequence Processing and Combinatorial Binding: Neurobiologically Royalsocietypublishing.Org/Journal/Rstb and Computationally Informed Hypotheses

Structured sequence processing and combinatorial binding: neurobiologically royalsocietypublishing.org/journal/rstb and computationally informed hypotheses Ryan Calmus, Benjamin Wilson, Yukiko Kikuchi and Christopher I. Petkov Review Newcastle University Medical School, Framlington Place, Newcastle upon Tyne, UK RC, 0000-0002-7826-9355; BW, 0000-0003-2043-7771; YK, 0000-0002-0365-0397; Cite this article: Calmus R, Wilson B, Kikuchi CIP, 0000-0002-4932-7907 Y, Petkov CI. 2019 Structured sequence processing and combinatorial binding: Understanding how the brain forms representations of structured information distributed in time is a challenging endeavour for the neuroscientific neurobiologically and computationally community, requiring computationally and neurobiologically informed informed hypotheses. Phil. Trans. R. Soc. B approaches. The neural mechanisms for segmenting continuous streams of sen- 375: 20190304. sory input and establishing representations of dependencies remain largely http://dx.doi.org/10.1098/rstb.2019.0304 unknown, as do the transformations and computations occurring between the brain regions involved in these aspects of sequence processing. We propose a blueprint for a neurobiologically informed and informing computational Accepted: 4 November 2019 model of sequence processing (entitled: Vector-symbolic Sequencing of Binding INstantiating Dependencies, or VS-BIND). This model is designed to support One contribution of 16 to a theme issue the transformation of serially ordered elements in sensory sequences into struc- ‘Towards mechanistic models of meaning tured representations of bound dependencies, readily operates on multiple composition’. timescales, and encodes or decodes sequences with respect to chunked items wherever dependencies occur in time. The model integrates established vector symbolic additive and conjunctive binding operators with neurobiologically Subject Areas: plausible oscillatory dynamics, and is compatible with modern spiking neural neuroscience, cognition, computational biology network simulation methods. We show that the model is capable of simulating previous findings from structured sequence processing tasksthat engage fronto- temporal regions, specifying mechanistic roles for regions such as prefrontal Keywords: areas 44/45 and the frontal operculum during interactions with sensory rep- computational modelling, sequence learning, resentations in temporal cortex. Finally, we are able to make predictions based serial order, chunking, language, binding on the configuration of the model alone that underscore the importance of serial position information, which requires input from time-sensitive cells, Authors for correspondence: known to reside in the hippocampus and dorsolateral prefrontal cortex. This article is part of the theme issue ‘Towards mechanistic models of Ryan Calmus meaning composition’. e-mail: [email protected] Christopher I. Petkov e-mail: [email protected] 1. Introduction Natural environments are richly structured in both space and time. Substantial progress has been made in understanding the neurobiological bases of learned relationships between spatially or temporally separated elements [1–3]. More- over, prior research has established the importance of serial order for the brain [4], and binding problems, whereby distinct sensory events are combined for perception, decision and action [5], have attracted considerable interest and empirical enquiry [6,7]. Establishing relationships, or dependencies, between elements over time allows us to extract the structure of the sensory world and to make predictions about future events. However, understanding how the brain binds complex information distributed in time, building temporally organized structures that represent multiple linked dependencies, remains a considerable challenge facing the neuroscientific community. This sort of cognitive structure-building is challenging for the brain to achieve because complex input must be discretized in time, resul- tant discrete items chunked and stored in memory, dependencies identified and © 2019 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. related items bound for perception (or other purposes), and rep- morphemes, words and phrases in sentences. The language 2 resentations of multiple dependencies maintained concurrently binding problem features the rapid detection of lexical symbols royalsocietypublishing.org/journal/rstb in memory to be further manipulated [8,9]. and the encoding of complex syntactic regularities at multiple Human language—written, spoken or signed—is a salient scales and temporal granularities. example of the complexity of the binding problem, because A large volume of neurobiological data implicate a it features syntactically organized dependencies between fronto-temporal brain system in various aspects of language semantic units [10]. Yet the problem of building complex processing [11,24,25]. Neurobiological signals associated with representations is also relevant to complex action sequences the chunking and parsing of speech with respect to phrase [8,11,12], music [8,13,14], mathematics [9,15] and cognition in boundaries have been identified in these regions [26–28]. More- general [11,13]. Moreover, some of these systems are not over, the temporal structure of speech or language content at unique to humans, since songbirds can construct complex different timescales (phonemic, syllabic, word or phrasal) pro- vocalization sequences [16], an ability supported by a forebrain duces stimulus- or context-driven neural entrainment at the neural system [17], and correspondences have been established relevant oscillatory frequency bands [29,30]. between humans and a number of species in processing adja- Behaviourally, a number of sequencing processes are now Phil. Trans. R. Soc. B cent and non-adjacent sequencing dependencies [18–20]. known to be evolutionarily conserved, including entrainment Thus, advancing our understanding of how complex structure to rhythmic sensory input [29,31]. There is also information can be built from sequential input in computational and from artificial grammar learning paradigms, which are used neural systems is important for developing better machine to establish dependencies between otherwise arbitrary audi- and animal models to understand both general principles tory or visual items in a sequence, either via statistical or and species-specific aspects of combinatorial binding. rule learning [32,33]. Humans and a number of non-human 375 In this article, we propose a blueprint for a neurobio- animals can learn dependencies between sensory items next : 20190304 logically informed and informing computational model of to each other in a sequence (adjacent dependencies), as well sequence processing (entitled: Vector-symbolic Sequencing as dependencies further separated in time and by intervening of Binding INstantiating Dependencies, or VS-BIND). The items (non-adjacent dependencies; reviewed in [18]). The learn- VS-BIND approach integrates: (1) advances in modelling ing of hierarchically organized dependencies by non-human combinatorial binding within simulated neural systems using animals is, however, contentious and it remains to be seen vector symbolic operations; (2) insights from neuroimaging whether this ability is uniquely human [34,35]. and neurophysiological evidence in human and non-human Comparative neuroimaging work has identified brain primates on neural correlates of structured sequence proces- regions in human and monkey frontal and temporal cortex sing and working memory; and (3) dynamic mechanisms for involved in processing sequencing dependencies [19,20]. manipulating population codes that can be incorporated This has led to descriptive models of the brain-bases of struc- in modern spiking neural networks [21,22]. The approach tured sequence processing and the relationship with, and allows us to plausibly transform internal representations, ren- distinctions from, neurobiological processes involved in dering these into both mathematically idealized and neurally language [34,36,37]. For the purposes of this paper, we will simulated site-specific activity unfolding over time. Building focus on the Wilson et al. [34] descriptive neurobiological on these foundational mechanisms, we focus on modelling model of human and non-human primate structured chunk encoding and the binding of sensory items to represent sequence processing, shown in figure 1. Relatively simple adjacent, non-adjacent and more complex (hierarchically) sequencing relationships, such as between items that occur structured sequencing dependencies. Our key objectives here next to each other in a sequence, can be learned by both are to motivate the approach, ground it in the relevant litera- humans and monkeys and are seen to engage corresponding ture, and use it to generate distinct mechanistic predictions, brain regions, particularly the ventral frontal and opercular the form of which, as will be seen below, depends on both cortex (figure 1a, vFOC, bottom row) [34]. For non-adjacent the specific binding operations used and their configuration. dependencies (which increase working memory demands: It is important to note that any model using combinatorial store an item in memory long enough to link to its matched operators can only be described as classically

View Full Text

Details

  • File Type
    pdf
  • Upload Time
    -
  • Content Languages
    English
  • Upload User
    Anonymous/Not logged-in
  • File Pages
    13 Page
  • File Size
    -

Download

Channel Download Status
Express Download Enable

Copyright

We respect the copyrights and intellectual property rights of all users. All uploaded documents are either original works of the uploader or authorized works of the rightful owners.

  • Not to be reproduced or distributed without explicit permission.
  • Not used for commercial purposes outside of approved use cases.
  • Not used to infringe on the rights of the original creators.
  • If you believe any content infringes your copyright, please contact us immediately.

Support

For help with questions, suggestions, or problems, please contact us