Cross-frequency phase coupling of brain oscillations and relevance attribution as saliency detection in abstract reasoning

Highlights

• Salience attribution showed specific functional connectivity in abstract reasoning.

• Conventional, aberrant and futile relevance detection models were investigated.

• Theta-alpha phase synchrony was observed at frontotemporal, central, parietal sites.

• Strength of interactions and their topography were different over the models.

• Salience attribution in healthy participants can be phenotyped using phase synchrony.

Abstract

Abstract reasoning is associated with the ability to detect relations among objects, ideas, events. It underlies the understanding of other individuals’ thoughts and intentions. In natural settings, individuals have to infer relevant associations that have proven to be reliable or precise predictors. Salience theory suggests that the attribution of meaning to stimulus depends on their contingency, saliency, and relevance to adaptation. So far, subjective estimates of relevance have mostly been explored in motivation and implicit learning. Mechanisms underlying formation of associations in abstract thinking with regard to their subjective relevance, or salience, are not clear. Applying novel computational methods, we investigated relevance detection in categorization tasks in 17 healthy individuals. Two models of relevance detection were developed: a conventional one with nouns from the same semantic category, an aberrant one based on an insignificant common feature. Control condition introduced non-related words. The participants were to detect either a relevant principle or an insignificant feature to group presented words. In control condition they inferred that the stimuli were irrelevant to any grouping idea. Cross-frequency phase coupling analysis revealed statistically distinct patterns of synchronization representing search and decision in the models of normal and aberrant relevance detection. Significantly distinct frontotemporal functional networks with central and parietal components in the theta and alpha frequency bands may reflect differences in relevance detection.

Introduction

The ability to detect relations among objects, ideas, or situations is believed to be a precondition of higher order thinking (Dumontheil, 2014). This ability is related to abstract reasoning and probably underlies adaptive behaviour which is based on the generalization of rules to a novel environment (Badre et al., 2010; Kayser and D’Esposito, 2013). Abstract reasoning refers to manipulating relations among representations rather than stimulus features (Dumontheil, 2014). In natural settings, numerous associations are potentially important and individuals have to infer relevant ones that have proven to be reliable or precise predictors. An association is accurate and reliable when it yields general relevant knowledge which allows to reliably predict specific events. Alternatively, irrelevant associations that seem to be redundant and unreliable should be ignored.

Schizophrenic patients exhibit information‐processing abnormalities connected with a disability to discriminate relevant stimuli in a range of relevant and irrelevant ones (Nuechterlein et al., 1994; Orosz et al., 2011; Howes and Nour, 2016). According to the aberrant salience hypothesis, increased attribution of meaning (salience) to irrelevant events and reduced learning of relevant information may relate to an underlying deficit in relevance detection (Jensen et al., 2008). This salience hypothesis needs an objective experimental measurement of salience and implies the subjective nature of relevance and salience attribution (Katthagen et al., 2018). The subjectivity aspect can be approached by modelling unsupervised process of formation of associations which is a common way of describing unsupervised learning in categorization tasks (Coleravy and Lewandowsky, 2008).

Although there has been successful attempts to investigating learning from relevant and irrelevant stimuli (Katthagen et al., 2018), the role of salience attribution in abstract reasoning has not been addressed to. Nor have been investigated neuronal interactions with regard to the process of relevance detection in abstract reasoning. Meanwhile, this line of research could advance our knowledge of both salience attribution and higher order thinking.

Both constructs, relevance and salience are very close. In general, a feature can be considered subjectively salient when it is relevant in a particular context (Katthagen et al., 2018). Here, we define salient features as those that are used to infer a grouping principle in categorization tasks. In conceptual knowledge relations connecting representations are either categorical or thematic (Goswami, 2008). For instance, pear, apple and banana exemplify the idea of fruit. Thus, the relevant grouping idea would be the category of fruit. This example illustrates the model of normal (direct) relevance detection in our experiment. In schizophrenia insignificant features can turn subjectively relevant, or salient, and, therefore, disrupt normal learning and reasoning (Kapur, 2003). As a result, context-inappropriate associations are reinforced (Jensen et al., 2008). Therefore, the model of aberrant relevance detection in abstract reasoning in our experiment is based on generalization driven by insignificant, or latent features, which are normally filtered out. Thus, blade, coin and hook can be grouped based on the idea of shiny which is not a typical (categorical or thematical) way to generalize otherwise these non-related nouns. Similar tests for detecting associations connecting words or objects were adopted as an additional diagnostic measure in schizophrenia (Nikolaeva, 2011; Zeigarnik, 1972).

The evaluation of neuronal interactions in EEG data is a promising methodology referring to the analysis of abstract reasoning and processes composing it. Cross-frequency phase coupling is probably the optimal level of description of neurocognitive processes, integrating their genetic, structural, neurochemical, and bioelectrical aspects (Chaieb et al., 2015). Phase coupling between neuronal oscillations allows to integrate distant populations of neurons (Palva, Palva, 2018) when various motor, perceptual and cognitive tasks are performed (Palva and Palva, 2011, 2018). Phase synchronization is regarded as the essential form of neuronal coupling since it probably enables spiking propagation between neuronal populations (Guetig, 2014) and coordinates the interactions of fast and slow oscillations (Siebenhühner et al., 2016). There is numerous evidence that phase interactions in different frequency bins are associated with various functional roles. Thus, interactions between oscillatory sources in the alpha (∼8–12 Hz) and in the theta (∼4–6 Hz) bands in frontocortical areas may support higher cognitive functions (Palva and Palva, 2011) including attentional processes (Sauseng and Klimesch, 2008). In the present study we have chosen theta-alpha interactions because they are specific to categorization tasks in abstract reasoning. In particular, there is evidence that increased activity in the alpha frequency range reflects access to knowledge systems including long-term memory (Klimesch, 2012). Pronounced theta activity is linked to controlled semantic processing (Marko et al., 2019). Increased theta phase-locking in language areas (temporal lobe) was associated with semantic activation (Salisbury and Taylor, 2012). Also, theta locked activity is associated with cognitive control (Cavanagh and Frank, 2014).

Although there has been multiple evidence of involvement of theta and alpha rhythmic oscillations in cognitive processes composing higher cognitive functions, there is little research into associated EEG functional connectivity in abstract thinking and in saliency attribution underlying the formation of associations. Therefore, in the present study the investigation of theta–alpha phase interactions was of primary interest since these interactions probably reflect relevance detection in abstract reasoning.

The aim of the current study was to investigate saliency attribution as a basis for the formation of associations in abstract reasoning and associated neuronal signatures.

In our study we were looking for neuronal signatures of saliency detection over the course of formation of associations in the models of normal and aberrant saliency attribution. Normal saliency attribution implies that presented stimuli belong to one and the same category. The model of aberrant saliency attribution suggests that grouping is based on a latent, insignificant attribute for otherwise non-related words. As a control condition we introduced randomly chosen non-related words. The process of solving categorization tasks was unsupervised, i.e. participants were not provided feedback with regard to their decision in order to pertain the subjectivity aspect. We removed the EEG data from further analyses when a participant misclassified the stimuli in order to ensure the consistency of data across the participants. We hypothesised that the three models would be accompanied with statistically distinct neuronal signatures, reflecting normal, aberrant saliency attribution and a failure to detect a grouping principle in categorization tasks. Therefore, we first obtained EEG functional connectivity estimates in each participant separately and conducted group-level statistical analyses of those estimates afterwards. Based on previous research into processes composing abstract reasoning we looked at cross frequency phase interactions in the theta and alpha frequency bins.

In the current study 17 healthy participants performed categorization tasks while EEG was being recorded. In these tasks, participants had to discriminate between three types of association among the presented words via button press. We designed three conditions when either a categorical grouping principle (D) or an insignificant, latent attribute underlying presented words (L) was present, and when there was no relation between presented words (N). The participants had to figure out which type of association (D, L, N) was implied. We applied the general cross-frequency decomposition method (GCFD) to estimate phase interactions in the theta and alpha frequency bins accompanying the process of formation of associations. We hypothesised that topographies and phase locking values of obtained synchronous signals would be statistically different with regard to each condition.