PIs: Michael F. Bonner & Brice Ménard (Krieger School of Arts & Sciences)
The human brain is one of the most complex objects in the universe, and the fundamental principles of how it supports intelligent behavior remain unknown. One challenging aspect is that its functions are carried out in high dimensions— they rely on the coordinated interplay of immense highdimensional populations of neurons. All high-dimensional systems, including the brain, present several challenges to researchers: 1) they are difficult to observe and measure at scale, 2) analyzing data from high-dimensional systems requires major computational resources and advanced statistical methods, and 3) the intuitions we obtain from lower-dimensional systems often fail to generalize to higher dimensions. We believe that these challenges have hampered progress toward a neuroscientific theory of natural intelligence and that, with a new statistical framework and large-scale neuroscience data, we can transform how neuroscientists approach the study of the human brain and its high-dimensional functions.
Neuroscientists have struggled to understand the brain in part because the standard methodological tools of neuroscience are ill-equipped for characterizing statistical phenomena in high dimensions. In fact, the methods of neuroscience primarily focus on low-dimensional problems in small-scale datasets (e.g., identifying whether a specific brain region or neuron responds more to stimuli from category A or category B). As a result, there is a critical gap in our understanding of how the human brain operates in high dimensions. The goal of this proposal is to bridge this knowledge gap by applying a novel statistical technique to massive recordings of human brain activity and behavioral assessments of cognitive abilities. Specifically, we will analyze a publicly available large-scale dataset to test the hypothesis that our statistical approach can reveal new aspects of high-dimensional human brain representations that are inaccessible with conventional methods but may be critical for understanding individual differences in cognition.
The findings from this project have broad implications for understanding the fundamental properties of brain representation that vary across people and underlie differences in sensory and cognitive abilities. This project also has the potential to transform how neuroscientists study individual differences in brain function—opening new opportunities to explore high-dimensional representations and the precise ways in which they are shared or unique across people. Importantly, our findings will provide critical preliminary evidence for a proposal that we are currently developing to study the neural underpinnings of visual intelligence.
We propose to conduct a proof-of-principle analysis showing that individual differences in general intelligence are linked to high-dimensional aspects of human brain representations that would otherwise be undetectable with standard analysis methods. We will do so using a publicly available NIH-funded big-data initiative known as the Human Connectome Project (Van Essen et al., 2013). The Human Connectome Project contains one of the largest datasets in the world for functional neuroimaging of the human brain, and it includes rich behavioral and demographic data, allowing us to precisely characterize the relationship between the functional properties of the human brain and cognitive performance. The PIs, Bonner and Ménard, have recently developed a new statistical approach for characterizing a key functional signal in brain activity that spans thousands of orthogonal dimensions (Gauthaman et al., in prep). In contrast, standard analysis methods only allow researchers to see the tip of the iceberg of meaningful dimensions in human brain activity. We are now seeking to leverage these methodological innovations to make major advances in understanding the nature of diversity in human brains and behavior.
Our approach, which we refer to as crossdecomposition, involves two key procedures that allow us to uncover a high-dimensional spectrum of meaningful signals in the brain:
Hyperalignment: The fundamental dimensions of brain representation are not neurons or cortical patches but rather latent dimensions that need to be aligned across individuals. Our procedure optimally rotates the highdimensional brain representations of one individual to be functionally aligned with another. By doing so, we can uncover fine-scale information in human brain representations that is undetectable with conventional methods relying on coarse anatomical alignment.
Cross-validated dimensionality spectrum: When using standard methods to characterize the latent dimensions of brain representations, it is impossible to distinguish reliable signal from sources of nuisance variance, such as physiological noise and equipment noise. Our procedure uses cross-validation across repeated presentations of experimental stimuli to characterize dimensionality without corruption from nuisance variance, giving us the first unbiased view of the full spectrum of latent dimensions in human brain representations.
Using this approach, we will characterize the highdimensional spectrum of brain representations measured with functional neuroimaging, and we will test the prediction that newly detectable fine-scale information in brain activity is tightly connected to individual differences in fluid intelligence, which is the general ability to think abstractly and reason through problems.
