Feature-selective Attention in Frontoparietal Cortex: Multivoxel Codes Adjust to Prioritize Task-relevant Information

Abstract

Human cognition is characterized by astounding flexibility, enabling us to select appropriate information according to the objectives of our current task. A circuit of frontal and parietal brain regions, often referred to as the frontoparietal attention network or multiple-demand (MD) regions, are believed to play a fundamental role in this flexibility. There is evidence that these regions dynamically adjust their responses to selectively process information that is currently relevant for behavior, as proposed by the "adaptive coding hypothesis" [Duncan, J. An adaptive coding model of neural function in prefrontal cortex. Nature Reviews Neuroscience, 2, 820-829, 2001]. Could this provide a neural mechanism for feature-selective attention, the process by which we preferentially process one feature of a stimulus over another? We used multivariate pattern analysis of fMRI data during a perceptually challenging categorization task to investigate whether the representation of visual object features in the MD regions flexibly adjusts according to task relevance. Participants were trained to categorize visually similar novel objects along two orthogonal stimulus dimensions (length/orientation) and performed short alternating blocks in which only one of these dimensions was relevant. We found that multivoxel patterns of activation in the MD regions encoded the task-relevant distinctions more strongly than the task-irrelevant distinctions: The MD regions discriminated between stimuli of different lengths when length was relevant and between the same objects according to orientation when orientation was relevant. The data suggest a flexible neural system that adjusts its representation of visual objects to preferentially encode stimulus features that are currently relevant for behavior, providing a neural mechanism for feature-selective attention.

Spatial and feature-selective attention have distinct effects on population-level tuning

“…Our finding that prefrontal cortex appears to shape responses in occipital areas is consistent with work demonstrating that the responses of frontoparietal regions contain stimulus-related information (for example, [64]), that increases with spatial [15] and feature-selective [58] attention. One prominent model of prefrontal cortex function (biased competition model [41,65]) proposes that the prefrontal cortex biases processing in more specialized (visual) cortices in favor of task-relevant information.…”

“…Previous neuroimaging work has revealed some of the e↵ects of spatial [54,55,14,56] and feature-selective [9,10,11,12,57,58] attention at a population level. Like some previous fMRI studies, we used classifier accuracy as an intuitive means of measuring the e↵ects of attention: using classifier accuracy as a proxy for the amount of information that is potentially available in the neural response.…”

“…We created novel object stimuli that varied in color and in their shape statistics (see Figure 1B), using custom code. The shapes were variants of 'spikie' stimuli used in previous work [76,15,58]. All our 'spikie' shapes had a common almost spherical body and 16 spikes varying in location, length and orientation.…”

Abstract neural representations of category membership beyond information coding stimulus or response

“…Prefrontal cortex is strongly implicated representing abstract task variables (Duncan, 2001;Miller & Cohen, 2001) and task-relevant sensory signals (e.g. Erez & Duncan, 2015;Goldman-Rakic, 1995;Jackson et al, 2017;Meyers et al, 2008;Roy et al, 2010). We selected prefrontal regions implicated in cognitive control and contain taskrelated representations from the multiple-demand system (Duncan, 2010;Fedorenko et al, 2013)http://imaging.mrc-cbu.cam.ac.uk/imaging/MDsystem) including the posterior, middle (approximately area 8), and anterior (approximately area 9) portion of the middle frontal gyrus (MFG).…”

Concurrent neuroimaging and neurostimulation reveals a causal role for dlPFC in coding of task-relevant information

“…In the human brain, a network referred to as the multiple-demand (MD) network encodes a range of task features (Woolgar, Jackson, & Duncan, 2016) with a strong preference for attended information over information that is irrelevant (Erez & Duncan, 2015;Woolgar, Williams, & Rich, 2015;Jackson, Rich, Williams, & Woolgar, 2017;Jackson & Woolgar, 2018). This network includes the dorsolateral prefrontal cortex (dlPFC), the anterior insula and frontal operculum (AI/FO), intraparietal sulcus (IPS), and the pre-supplementary motor area and adjacent anterior cingulate (ACC/pre-SMA).…”