Neural representation dynamics reveal computational principles of cognitive task learning

941 

Abstract Submission 

Ravi Mill¹, Michael Cole¹

¹Rutgers University-Newark, Newark, NJ

Ravi Mill  
Rutgers University-Newark
Newark, NJ

Michael Cole  
Rutgers University-Newark
Newark, NJ

Learning cognitive tasks is a ubiquitous feature of everyday life, yet the neural basis for this essential human skill remains unclear. During learning, neural task representations must be rapidly constructed for novel task performance, then optimized for robust practiced task performance. The present study interrogated changes to neural representational geometry that underpin this transition from novel to practiced task performance. We integrated computational ideas from multiple neuroscientific sub-fields (long-term memory, contextual decision-making and cognitive control) to test a task learning framework underpinned by subcortical-cortical representational dynamics (Figure 1). The core hypothesis of this framework is that practice involves a shift in the brain from compositional representations (task-general activity patterns that can be flexibly reused across tasks) to conjunctive representations (task-specific activity patterns specialized for the current task).

Multiband functional MRI (fMRI) data were recorded from 44 healthy young adults (age mean=22.25, age range=18-36; 23 female) as they performed 2 sessions of our newly developed concrete permuted rule operations (C-PRO2) paradigm. This allows for the presentation of 64 distinct multi-sensory tasks, based on permutations of individual logic, sensory and motor rules (Figure 2). In the first session (Practice), subjects practiced a subset of 4 tasks repeatedly. In the second session (Test), these practiced tasks were intermixed with the remaining 60 novel tasks. The design therefore enabled fMRI recording as multiple complex tasks were performed from first novel presentation through repeated practice. The resulting fMRI data underwent preprocessing (Glasser et al., 2013) and general linear modeling to estimate task activations for cortical vertices and subcortical voxels. A functional atlas (Glasser et al., 2016; Ji, Spronk et al., 2019) was used to affiliate these vertices/voxels to regions and large-scale functional networks, which was key for our multivariate analyses.

We firstly observed behavioral evidence of learning, in the form of improvements to accuracy and reaction time as tasks were repeatedly practiced. We then developed analytic methods inspired by multivariate pattern analysis and machine learning to interrogate changes to neural representational geometry that underpinned these behavioral practice effects. Specifically, novel tasks presented in the held-out Test session were used to build neural templates of individual (i.e. compositional) rule types, and their non-linear interaction (i.e. conjunction). These were fit simultaneously via multiple linear regression to each of the practiced tasks in the Practice session, so as to dynamically quantify the strength of rule compositions and rule conjunctions over task learning. The results substantiated the hypothesized dynamic shift from compositional to conjunctive representations with repeated task practice. Further, we found that conjunctions originated in subcortex (hippocampus and cerebellum) and slowly spread to cortex. Critically, it was the strengthening of conjunctive representations in cortex that was uniquely associated with signatures of effective task practice (improved behavior and reduced task interference).

Our findings reveal a precise neural mechanism underlying changes to neural representational geometry that occur over learning: increasing non-linear conjunction (binding) of task rule elements. This extends computational concepts from long-term memory (O'Reilly & Rudy, 2001) and contextual decision-making (Kikumoto & Mayr, 2020) to the domain of cognitive task learning. The formation of cortical conjunctive representations hence serves as a computational mechanism of effective task practice, reflecting cortical-subcortical dynamics that optimize task representations in the human brain.

Executive Function, Cognitive Control and Decision Making ¹

Learning and Memory Other ²

Cognition

Cortex

Experimental Design

FUNCTIONAL MRI

Learning

Machine Learning

Memory

Motor

Sub-Cortical