Scientific Understanding of Consciousness
Consciousness as an Emergent Property of Thalamocortical Activity

Brain Neuroscience Methods


Nature  526, 371–379 (15 October 2015)

Progress and challenges in probing the human brain

Russell A. Poldrack & Martha J. Farah

Department of Psychology, Stanford University, Stanford, California 94305, USA

Center for Neuroscience & Society, University of Pennsylvania, Philadelphia, Pennsylvania 19104, USA


Perhaps one of the greatest scientific challenges is to understand the human brain. Here we review current methods in human neuroscience, highlighting the ways that they have been used to study the neural bases of the human mind. We begin with a consideration of different levels of description relevant to human neuroscience, from molecules to large-scale networks, and then review the methods that probe these levels and the ability of these methods to test hypotheses about causal mechanisms. Functional MRI is considered in particular detail, as it has been responsible for much of the recent growth of human neuroscience research. We briefly review its inferential strengths and weaknesses and present examples of new analytic approaches that allow inferences beyond simple localization of psychological processes. Finally, we review the prospects for real-world applications and new scientific challenges for human neuroscience.

The way that we conceptualize brain function has always been constrained by the methods available to study it. Studies of patients with focal brain lesions in the nineteenth century led to the view of the brain as a collection of focal centres specialized for particular cognitive abilties, such as ‘Broca’s area’ for speech production. The development of neurophysiological recording techniques in the twentieth century led to Barlow’s ‘neuron doctrine’, according to which the functions of individual neurons can be extrapolated to explain the function of the brain as a whole. The cognitive neuroimaging studies of the 1980s focused on subtractive comparisons between cognitive tasks meant to isolate specific cognitive operations, and led to a relatively modular view of brain function as involving localized and separable regions that implement elementary mental operations.

The methods of contemporary human neuroscience have provided a much more complex and nuanced view of the human brain as a dynamic network with multiple levels of organization, in which function is characterized by a balance of regional specialization and network integration. Although current methods are limited in their utility for studying brain function at fine-grained levels of organization (such as single neurons or cortical columns), human neuroscience has nonetheless made remarkable progress in understanding basic aspects of functional organization, and with this have come a number of applications to address real-world problems. Our goal here is to review the current state of human neuroscience, focusing on what kinds of questions can and cannot be answered using current techniques and how those answers are relevant to real-world applications.

Methods for studying human brain function can be organized according to the kinds of mechanistic insights that each technique provides. Positron emission tomography (PET) using neurotransmitter ligands measures molecular mechanisms, even though its spatial resolution is on the order of one centimetre. The second characteristic is the ability of each method to elucidate the mechanistic role of an observed brain molecule, cell, region or network in a mental function of interest. By mechanism we mean the causal chain of events that result in the realization of a function. To fully understand human brain function is to know the causal chains of events at the molecular, cellular, population, and network levels that give rise to psychological function. For this reason, the power to identify causal relationships is a crucial dimension of difference among methods.

Associations provide a valuable starting point for theorizing about the neural mechanisms of human psychology, and their evidentiary value can be strengthened by measuring possible confounds to rule them in or out.

Although functional neuroimaging, electroencelphalography/magnetoencelphalography (EEG/MEG) and single-cell recordings are sometimes criticized as being purely correlative and therefore uninformative about mechanism, that criticism is only partly accurate. When psychological processes are experimentally manipulated by presenting a certain kind of stimulus and/or engaging the subject in a task, we can infer that any reliably elicited brain activity was caused by performing these psychological functions. We cannot, however, infer with confidence that the observed brain activity is causally responsible for the psychological process under study. Despite this limitation (which is shared by neuronal recordings in non-human animals), neuroimaging studies in which psychological processes are manipulated comprise the majority of current human neuroscience research, and have advanced our understanding of human brain function,

Use of non-invasive brain stimulation for research purposes has grown rapidly in recent decades, starting with transcranial magnetic stimulation (TMS), in which pulsed magnetic fields induce currents in the brain. Various forms of transcranial electric stimulation (TES), in which current is delivered using external electrodes, have also been used, of which the most common variant is transcranial direct current stimulation (tDCS). Unlike DBS, non-invasive brain stimulation generally affects larger and more superficial areas of the brain, but researchers are seeking to improve spatial resolution with new magnetic coil shapes for TMS and new electrode configurations for tDCS. Focused ultrasound is also being explored as a means to stimulate more precisely delimited brain regions. Pharmacological agonists and antagonists of particular neurotransmitter systems can be used to experimentally manipulate the human brain at the molecular level, although with imperfect specificity. By combining each of these manipulations of brain function with functional brain imaging, one can leverage the causal information obtained through pharmacological challenges or brain stimulation. For example, the causal role of activity in specific brain regions, identified using fMRI, for a particular function has been tested by brain stimulation, using both direct cortical stimulation and TMS.

Because fMRI has become the main method for the study of human brain function, our review focuses on this method and new ways of using it. In the last two decades, fMRI has transitioned from a newly developed technique for revealing neuronal activity to being the workhorse method of cognitive neuroscience Much has been learned about the biological mechanisms underlying blood oxygen level dependent (BOLD) signals, but still much remains to be understood, such as the roles of specific glial and neuronal cell types in the coupling of neuronal activity to blood flow. This limited physiological understanding poses problems for the interpretion of fMRI data. In particular, although fMRI signals often correlate strongly with both action potentials (‘spikes’) and local field potentials, they are largely reflective of post-synaptic processes, and in some cases they can be dissociated from spiking altogether. The relative sensitivity of fMRI to post-synaptic processes as opposed to spiking has been seen as a drawback by some who view spikes as the essence of brain function, but it is worth noting that this discovery has actually rekindled interest in the analysis of local field potentials in electrophysiology (where these signals have long been discarded) and suggests that fMRI may sometimes be sensitive to subthreshold signals that would be missed by analysis of spikes only. Uncertainties in relating fMRI to psychological, as well as physiological, processes have also been debated, and progress has been made on this front too. From experimental approaches such as adaptation paradigms for probing representations to analyses of functional connectivity, fMRI is routinely used to answer questions about mind–brain relationships that go far beyond localization. Here we discuss three examples of new approaches to understanding human brain function with fMRI that address questions of representation, computational processes and network interactions across the brain.

Computational models play a central role in our understanding of both cognitive and brain functions and, increasingly, of the relationship between the two. By making assumptions explicit, computational models enable more direct testing of theories, as well as providing the means to link computations at the neuronal level with higher-order functions. An example of an area in which substantial progress has been made using this approach is reinforcement learning, in which an animal selects actions and learns from the rewards gained from those actions. Computational models of reinforcement learning (RL) have long played a central role in artificial intelligence and psychology, and the discovery by Schultz and colleagues that dopamine neurons appear to signal one of the important quantities in these models (reward prediction error) has brought these models to the forefront of the neuroscience of decision making. For example, a set of publications in 2003 applied RL models to neuroimaging data and thereby identified correlates of reward prediction error signals in dopaminergic target regions such as the ventral striatum. Subsequent neuroimaging work has established that there are multiple RL signals in the brain, some reflecting the simple association between actions and values (known as ‘model-free’ RL) and others reflecting more complex contextual and hierarchical learning processes (known as ‘model-based’ RL. Similarly, in the study of memory, progress has been made in the mapping of medial temporal lobe subregions to specific computational operations such as pattern completion and pattern separation. In each of these domains, the computational interpretation of neuroimaging signals has been greatly enhanced by parallel studies in non-human animals, allowing imaging signals to be linked more directly to direct measures of neuronal activity.

New technologies for imaging and manipulating the human brain

Rapid advances in non-human neuroscience have been driven by the development of technologies that measure and manipulate brain function with increasing precision. Human neuroscience has lagged in this respect, in part because of the ethical challenges associated with direct manipulation and neuronal recording of the human brain. However, in response to the urgent need for better treatments for psychiatric disorders, research is underway with the aim to design implantable systems for sensing and modulating human brain networks. The development of optogenetic and ‘opto-fMRI’ approaches in non-human primates suggests that these methods may one day become feasible for use in human studies, and it is likely that electrical brain stimulation will eventually be supplemented with optogenetic approaches. Although such invasive techniques will likely only be used in rare clinical cases (that is, patients are undergoing implantation for medical reasons), they have the potential to provide much greater specificity in circuit mapping.

fMRI will probably remain the principal neuroimaging method in humans in the foreseeable future. However, the ongoing BRAIN initiative in the United States is providing substantial funding to develop entirely new techniques for imaging of brain function, and a significant proportion of this funding will go specifically towards the development of new methods for imaging the human brain. In addition, new developments in MRI have greatly increased the utility of standard MRI systems. For example, multiband imaging techniques have enabled a several-fold increase in the temporal resolution of fMRI acquisitions, and higher MRI field strengths (7 tesla and higher) hold promise to enable improvements in spatial resolution as well. There is thus great reason to be optimistic that methodological limits will continue to be pushed in the future.

Additional insight into human brain function will likely come from the study of postmortem human brains, which has long been a staple method for the characterization of anatomical structure and study of brain disorders. New techniques have enhanced the ability to visualize the structure of human brain tissue. For example, optical coherence tomography has been used to image ex vivo human cortical tissue, providing high-resolution imaging of cytoarchitecture with less distortion than standard microscopy techniques. The first whole-brain atlas of genome-wide gene expression in postmortem human brains has provided an important resource for understanding how gene expression relates to brain function; for example, the maps from this project have been used to identify expression differences across different resting-state networks. Continued development of such resources will be essential for progress in understanding the genetic architecture of brain function and their relation to mental health disorders.


The Human Connectome Project is nearing completion, and has already provided a rich database for the modelling of functional and anatomical connectivity of the human brain. However, fundamental challenges remain. For example, diffusion MRI provides the means to track white matter pathways and has been used to identify white matter connectivity disruptions associated with cognitive disorders such as dyslexia; however, diffusion imaging has inherent biases that limit its ability to accurately track connections across the entire brain. The last decade has seen a proliferation of approaches to model functional connectivity on the basis of functional MRI data, though the dust has yet to settle regarding which methods are most effective. To determine this, the analysis methods must be validated, which is challenging to do in humans but may be achieved using direct measurements of functional connectivity from invasive human approaches and non-human animals to validate the neuroimaging results. There is increasing evidence that at least in non-human primates functional connectivity reflects anatomical connectivity as measured using either diffusion MRI81 or anatomical tract-tracing; but it remains an important challenge to establish the ways in which functional and diffusion connectivity measures converge or diverge.

Reproducibility of neuroimaging research

Large-scale meta-analyses have made it clear that neuroimaging results can be highly convergent across studies, to the degree that cognitive processes can be accurately inferred from individual subject data using decoders trained on meta-analytic data based on reported activation coordinates. However, the last few years have also seen increasing concern regarding the reproducibility of research findings in neuroscience, paralleling more general concerns about reproducibility of scientific results. These issues are particularly acute for neuroimaging given the high dimensionality of the data, relatively low statistical power of many studies, high degree of analytic flexibility in data analysis procedures, and potential for questionable research practices such as circular analysis procedures. The field of neuroimaging has been at the forefront of a number of developments that aim to improve reproducibility and the sharing of data are increasingly being embraced. The Alzheimer’s Disease Neuroimaging Initiative (ADNI), International Neuroimaging Data Sharing Initiative (INDI), ENIGMA, and the Human Connectome Project together have shared thousands of neuroimaging data sets and this has enabled a number of novel discoveries. For example, data sharing by the ENIGMA consortium has enabled the first well-powered genome-wide association study of brain volume, identifying replicated associations between brain volume and several common genetic variants. In addition, nearly all of the main software packages for neuroimaging data analysis are free and open source, providing transparency and reproducibility in data analysis across groups, and the publication of fully reproducible analysis workflows has begun. The increasing use of machine learning methods, with their focus on out-of-sample generalization rather than statistical significance, is also leading to a greater emphasis on achieving reproducibility.

[end of paraphrase]


Return to – Brain Imaging