Human daily life hinges on the intricate interplay of various cognitive functions, encompassing perception, decision-making, and action. Cognitive neuroscience strives to decipher the complete neural representations underlying these functions. Previous research has successfully characterized perceptual experiences using encoding models, primarily focusing on passive tasks like listening or viewing. However, a comprehensive quantitative model describing the cortical organization of multiple active cognitive processes has remained elusive. This study aims to address this gap by employing fMRI to measure brain activity during the performance of a wide array of cognitive tasks. By utilizing two voxel-wise encoding models – a sparse task-type model and a cognitive factor model employing continuous metadata – the researchers sought to map the relationships between different cognitive processes and their cortical representations. The expectation was that this approach would not only reveal the hierarchical organization of cognitive functions but also offer insights into the generalizability of the models to novel, unseen tasks. The significance of this work lies in its potential to provide a robust framework for understanding the comprehensive cortical organization of human cognition, moving beyond passive observation to encompass the complexity of active cognitive engagement.

Prior research in cognitive neuroscience has utilized voxel-wise encoding and decoding models to explore the cortical basis of cognitive processes. These models, employing multivariate regression analysis, have successfully linked brain activity to various factors such as visual features, object categories, sound features, and linguistic information. Studies have also mapped semantic representational spaces onto the cortex, revealing important categorical dimensions. However, these studies largely relied on data from passive tasks (listening or viewing), leaving the cortical representations underlying active cognitive processes largely unexplored. This study builds upon this prior work by extending the scope to active cognitive processes and using a more comprehensive and diverse set of tasks.

Six healthy participants underwent fMRI scans while performing 103 diverse naturalistic cognitive tasks. The tasks ranged from simple sensory detection to complex processes like language comprehension and calculation. The researchers employed two encoding models: 1. **Task-type model:** This model used binary labels to represent each task, allowing for the examination of representational relationships among tasks across the cortex. A representational similarity matrix (RSM) and hierarchical cluster analysis (HCA) were used to reveal task relationships and cluster them into groups based on cortical activation patterns. Metadata-based reverse inference, using the Neurosynth database, was employed to interpret the cognitive factors associated with each cluster. 2. **Cognitive factor model:** This model utilized a 715-dimensional continuous feature space derived from metadata references in the Neurosynth database. This allowed for the representation of tasks as continuous vectors in a cognitive space, enabling predictions of brain activity and decoding of tasks, even for novel, unseen tasks. This was done by creating a cognitive transform function (CTF) that mapped task-type features into cognitive factor features. For both models, regularized linear regression was used to fit the models. The researchers also performed principal component analysis (PCA) to visualize the cognitive space and its cortical mapping. The methodology included rigorous statistical analyses to assess the significance of the findings, accounting for multiple comparisons using false discovery rate (FDR) correction. Additional analyses, including the use of a hierarchical model (based on the HCA results) and control analyses with sensorimotor regressors, were conducted to refine the interpretation of the results.

The task-type model revealed a hierarchical organization of the 103 cognitive tasks, clustered into six main groups based on their cortical representations: visual, auditory, motor, language, introspection, and memory. Metadata-based reverse inference confirmed the intuitive interpretations of these clusters. For instance, the visual cluster showed strong correlations with terms like "visual" and "perceptual." The hierarchical model, incorporating this hierarchical structure, outperformed the basic task-type model in predicting brain activity. PCA further revealed a gradual shift in the cognitive space from perceptual to more complex cognitive processes and mapped this organization onto the cortex. The cognitive factor model, using continuous metadata features, demonstrated impressive generalizability. It accurately predicted brain activity for novel tasks not included in the training data (mean prediction accuracy of 0.322 across the cortex) and achieved high accuracy in decoding novel tasks (mean decoding accuracy of 96.0%). Control analyses, regressing out low-level sensorimotor components, confirmed the model's ability to predict higher-order cognitive components. A detailed analysis of task selectivity in a specific brain region (left inferior parietal lobule) revealed a topographic organization, showcasing the model's ability to capture the intricate interplay of multiple cognitive dimensions within a single brain area.

This study offers a significant advancement in our understanding of the cortical organization of human cognition. Unlike previous studies that relied on passive tasks or focused on a few pre-defined cognitive factors, this research utilized a large and diverse set of naturalistic tasks to comprehensively sample the cognitive space. The high prediction and decoding accuracy for novel tasks, particularly with the cognitive factor model, demonstrates the robustness and generalizability of the approach. The findings corroborate previous research identifying certain cognitive clusters and components but extend these findings by offering a more quantitative and comprehensive framework for understanding their relationships and cortical organization. The hierarchical structure identified in the task representations supports a hierarchical model of cognitive processing within the brain. The study also highlights the power of metadata-based reverse inference for interpreting the cognitive dimensions captured by the models. The use of both the task-type and cognitive factor models provides a powerful combination of methods for exploring the complex relationships between cognitive processes and brain activity.

This study presents a novel framework for understanding the organization of diverse cognitive functions in the brain using quantitative modeling approaches. The results demonstrate the hierarchical organization of cognitive tasks and the generalizability of a cognitive factor model to novel tasks. This comprehensive analysis of a large dataset provides a valuable foundation for future research into individual differences in cognitive function and the development of personalized cognitive interventions. Future work could expand the range of tasks included, exploring cognitive domains currently underrepresented (such as odor perception and social interaction), further refining the models, and investigating the neural mechanisms underlying the identified hierarchical structure.

While this study used a large number of tasks, it did not encompass the full spectrum of human cognitive abilities. Future studies could incorporate a wider range of tasks, including those involving social interaction, olfactory processing, and more complex real-world scenarios. The study also relied on fMRI, which has limitations in terms of temporal resolution. The interpretation of cognitive factors from Neurosynth metadata is also subject to the biases and limitations of the underlying neuroimaging literature. Finally, the findings are based on a relatively small sample size of six participants. Replication with a larger, more diverse sample is needed to further validate the results.