Adaptive coding in the human brain: Distinct object features are encoded by overlapping voxels in frontoparietal cortex

Abstract

Our ability to flexibly switch between different tasks is a key component of cognitive control. Non-human primate (NHP) studies (e.g., Freedman, Riesenhuber, Poggio, & Miller, 2001) have shown that prefrontal neurons are re-used across tasks, re-configuring their responses to code currently relevant information. In a similar vein, in the human brain, the "multiple demand" (MD) system is suggested to exert control by adjusting its responses, selectively processing information in line with our current goals (Duncan, 2010). However, whether the same or different resources (underlying neural populations) in the human brain are recruited to solve different tasks remains elusive. In the present study, we aimed to bridge the gap between the NHP and human literature by examining human functional imaging data at an intermediate level of resolution: quantifying the extent to which single voxels contributed to multiple neural codes. Participants alternated between two tasks requiring the selection of feature information from two distinct sets of objects. We examined whether neural codes for the relevant stimulus features in the two different tasks depended on the same or different voxels. In line with the electrophysiological literature, MD voxels were more likely to contribute to multiple neural codes than we predicted based on permutation tests. Comparatively, in the visual system the neural codes depended on distinct sets of voxels. Our data emphasise the flexibility of the MD regions to re-configure their responses and adaptively code relevant information across different tasks.

Frontoparietal action-oriented codes support novel instruction implementation

“…Thus, our results suggest, first, that prefrontal representations carry action-oriented information during instruction following. This is line with previous studies that propose a crucial role of the frontolateral cortex in the integration of stimulus and response information into a task set based on verbal instructions 12,32,33 , as well as in representing task rules 17,24 and goals 34 . In contrast, parietal cortices contained both declarative and procedural information of relevant S-Rs.…”

“…As a first step, we investigated which brain regions were predominantly involved in instruction prioritization. Our intuition was that prioritization would boost implementation signals and, as such, we expected a frontoparietal network to be particularly crucial, as it is usually involved in the implementation of novel task sets 11,14–17,24 . We thus established a set of a priori candidate regions that encompassed frontal (inferior and middle frontal gyri) and (inferior and superior) parietal cortices (see Fig.…”

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

“…Duncan & Owen, 2000;Fedorenko et al, 2013) and have been posited to play a fundamental role in executive function and cognitive control (Corbetta & Shulman, 2002;Cole & Schneider, 2007). Non-human primate research has shown that these regions preferentially code task-relevant information (Rao, Rainer, & Miller, 1997;Freedman et al, 2001;Cromer et al, 2010;Roy et al, 2010;Kadohisa et al, 2013), a result mirrored in human neuroimaging data (Woolgar et al, 2015;Jackson et al, 2017;Jackson & Woolgar, 2018). This selective representation may support preferential coding in other brain regions (Desimone & Duncan, 1995;Miller & Cohen, 2001), as indicated by previous work (Baldauf & Desimone, 2014;Goddard, Carlson, & Woolgar, 2019), yielding a plausible mechanism for a brain-wide focus on relevant information and a key component of cognitive control.…”

“…MD ROIs were defined using co-ordinates from a previous review of activity associated with a diverse set of cognitive demands (Duncan & Owen, 2000) using the kernel method described in Cusack, Mitchell, and Duncan (2010) as in our previous work (Woolgar, Hampshire, Thompson, & Duncan, 2011;Woolgar, Thompson, Bor, & Duncan, 2011;Woolgar et al, 2015;Jackson et al, 2017;Jackson & Woolgar, 2018). We opted to use a template definition of the MD network because recent work has indicated that there is no particular benefit to using individual subject functional localisers to define these regions for multivariate analysis (Shashidhara, Spronkers, & Erez, 2019), whereas the template approach made it easy to compile the data across the group.…”

“…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).…”