Multitask representations in the human cortex transform along a sensory-to-motor hierarchy

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Neuroscience

Background:

• Human cognition relies on distributed neural processes, but the underlying computational and functional architectures are not fully understood.
• Characterizing how the brain represents information across various cognitive tasks is essential for understanding brain function.

Purpose of the Study:

• To investigate the geometry and topography of multitask representations across the human cortex.
• To uncover the computational principles organizing these representations to support multitask cognition.

Main Methods:

• Utilized functional magnetic resonance imaging (fMRI) in individuals performing 26 diverse cognitive tasks.
• Measured representational similarity within regions and alignment between regions across tasks.
• Trained multilayer neural network models to transform visual-to-motor representations.

Main Results:

• Representational alignment showed a graded organization along the sensory-association-motor axis.
• Multitask representational dimensionality exhibited a compression-then-expansion pattern along this gradient.
• This compression-expansion organization emerged in neural network models trained with a rich regime, leading to hierarchically structured, noise-robust representations.

Conclusions:

• The human cortex organizes multitask representations with a compression-then-expansion principle.
• This organization, particularly when learned through a rich training regime, supports robust and hierarchical information processing for multitask cognition.