Alan Turing’s revolutionary ideas about computation helped launch the field of cognitive science. One of his major contributions to cognitive science was the idea of a Turing machine, a hypothetical contraption capable of carrying out any algorithm or mechanical procedure using input and output symbols. Could the human mind be considered a kind of organic Turing machine? Could a machine be created that simulates a human mind? How does the human mind process information, anyway? These were the questions that inspired researchers in what became cognitive science, the study of human cognition. Scientists drawn to these questions come from many different disciplines, including biology, psychology, computer science, linguistics, and anthropology. 

Today, cognitive science includes many branches based on these disciplines as well as interdisciplinary work. Cognitive neuroscience is the branch that looks at how the brain works, so as to inform and constrain theories about cognition. Since the mid-century “cognitive revolution”, these branches have radiated out and away from each other, but now new developments in cognitive neuroscience offer possibilities for bringing them closer together.

Cognitive neuroscience may have originated with 19th century physician Pierre Paul Broca, who observed that patients with left-hemisphere brain damage suffered language problems. This finding was among of the first evidence that certain cognitive functions require certain areas of the brain, in contrast to theories that suggested the brain was an undifferentiated mass of cells. Since Broca’s time, we have found more and more cases of brain areas being specialized for certain functions. Today, cognitive neuroscience uses information about the brain not only from studies of people with brain damage, but also from studies of healthy brains thanks to innovations in neuroimaging methods.

Neuroimaging and cognitive neuroscience

Neuroimaging is the tool most commonly associated with cognitive neuroscience, and refers to any means of visualizing the brain; the great advantage of neuroimaging is that it lets us look inside the skulls of living people. Functional neuroimaging, such as PET (positron emission tomography) or fMRI (functional magnetic resonance imaging), allows scientists to learn about activity in the brain, whereas structural neuroimaging allows for detailed visualization of the structure of the brain (for example, to compare the brains of healthy and brain damaged individuals).

Today, most neuroimaging studies are based on some form of magnetic resonance imaging (MRI). The “scanners” are just like the MRI machines you might find in a hospital (though sometimes more powerful). One of the great advantages of MR neuroimaging is that, unlike PET or X-rays, magnetic resonance does not expose research participants to potentially dangerous radiation, so MRI studies are considered relatively safe and non-invasive. MRI can also produce high resolution structural images on the scale of millimeters, and fMRI can give information about brain activity on the scale of seconds. The basis of the MR signal is the response of water-based hydrogen ions (protons) in various settings — depending on what is around them, a different signal is obtained. In this way, the different signals can then be used to distinguish between different types of tissue in the cranium (for example, skull vs. brain), or between oxygenated and de-oxygenated blood (a proxy for brain activity and the basis of fMRI). The signals obtained from MRI can be analyzed in many different ways giving us new and exciting information about the brain and brain activity.

Some critics of neuroimaging have argued that it provides nothing more than a new form of “phrenology” because of its linking of structures (brain areas) with functions (e.g., language, decision making, memory, etc.). Phrenology was the 19th century craze for identifying aspects of intelligence or personality based on bumps on various parts of the skull, spearheaded by Franz Gall (and found to be without scientific merit by modern standards). Early in the history of neuroimaging, it is true that many studies did not go far beyond finding the neural correlates of cognitive functions (that is, the parts of the brain associated with those functions). However, advances in neuroimaging methods have allowed scientists to go far beyond these early steps, and now a second or third generation of neuroimaging methods is allowing them to ask new kinds of questions about the relationships between mind and brain.

Functional and structural connectivity: Beyond modern phrenology

Although some brain regions are more active for some tasks than others, it is unlikely that any area of the brain acts in isolation, or that any cognitive task could be carried out by a single area of the brain. Instead, different areas of the brain act as interconnected nodes in a network. However, until recently it has been difficult to study relationships among different brain areas in a living human brain. Early neuroimaging studies were limited to examining structure-function correspondences. New advances in neuroimaging methods are now allowing scientists to start asking questions about connectivity in the brain.

In structural connectivity studies, MRI data are used to visualize the physical connections between different brain regions. The brain is made up of cells called neurons that communicate electrically by sending directional signals through thin extensions of the cells called axons. Most axons are covered by an insulating material called myelin that makes signals travel faster down the axon. Diffusion Tensor Imaging (DTI) takes advantage of the fact that water molecules — which are the basis of MRI — cannot diffuse across the myelin. From this simple fact (and a little matrix algebra), a map of white matter tracts — the “wiring” of the brain — can be constructed (Figure 1).

Remember that DTI is a structural, not functional imaging method: it only shows what connections are anatomically likely, not which might be more involved in a particular task. However, DTI can be (and has been) used to visualize differences in white matter tracts across development (brains of different ages), between the brains of healthy individuals and those with various disorders of the nervous system, and even before and after certain types of mental training. This means that scientists can now see how connections between different brain areas vary with age or with disease, and whether some types of interventions affect the connections between brain areas rather than causing an effect in one particular area of the brain. For example, we now know that adults have more myelinated white matter tracts than children, and we know more about the order and rate at which they develop [1]. We know that in some diseases, connectivity is affected [2]. We even know that reading practice can strengthen white matter connectivity in students with reading difficulties [3]. Atypical development in white matter connectivity may also be detected before behavioral symptoms of a disease appear, so, in the future, DTI might be part of an early screening system to help provide support for those at risk of disorders before they develop [4].

Figure 1. An example of white matter tracts mapped with diffusion tensor imaging (DTI), shown from multiple perspectives. A: top; B; back; C: side; D: combined data for imaged tracts. (Image credit: Aaron Filler, MD, PhD)

Advances in magnetic resonance neuroimaging have also provided new methods for studying functional connectivity. Possibly the most prominent among these is resting state functional connectivity MRI (rs-fcMRI). As the name suggests, this method is based on the resting state of the brain — that is, what your brain is doing when you’re not doing anything in particular. When the brain is in a resting state, there are natural fluctuations in activity across different brain regions. Some brain regions show synchronized activity with other brain regions; this suggests that those areas are working together. Brain areas that are “in sync” must have some connections to each other, either directly or indirectly. Resting state functional connectivity analysis searches through fMRI data of a resting brain to find brain areas that have patterns of activation most like other areas. In addition to comparing patterns of activation between different populations (e.g., healthy vs. patient, or different age groups), rs-fcMRI can be used to ask what brain areas might be working together, i.e., functionally connected to each other.

Multi-voxel pattern analysis: Mind reading?

Another new neuroimaging method on the block is not explicitly about connectivity, but still gets around the “modern phrenology” problem. Multi-voxel pattern analysis (MVPA) is a way of comparing the pattern of activation in the entire brain between different states. Like conventional fMRI, MVPA contrasts brain activity in one behavioral condition as opposed to another (for example, looking at a picture of a flower vs. looking at a picture of a bird). However, whereas fMRI focuses on changes in activity in only one brain region or a small set of brain regions, MVPA can take into account the overall pattern of activation in the entire brain (up to several millimeter resolution). Therefore, conditions that can be contrasted in MVPA can be much more similar to each other than those in fMRI.  In fact, by using MVPA scientists have been able to distinguish between the brain states of a person looking at a single picture that can be interpreted in different ways (for example, a picture of a duck-rabbit that can be interpreted as a duck or a rabbit). More recently, scientists have used MVPA to reconstruct images of YouTube videos being viewed by research participants [5]. These astounding demonstrations of the potential power of MVPA have lead to concerns about the ethics of “mind reading” using neuroimaging technologies. Could the CIA one day use neuroimaging to read the minds of suspected criminals rather than interrogating them? While this scenario is not impossible some time in the future, for now the extent of MVPA “mind reading” is limited to carefully controlled laboratory situations that are unlikely to be available to current or future interrogators. Still, for those bioethicists among our readers, it never hurts to have the ethical considerations worked out before the technology is available (if ever)!

From neuroimaging to neural networks?

The questions raised by Turing’s ideas inspired many computer scientists to try and simulate human cognition with computers. This line of research can be roughly split into two types: research seeking to create machines as adept as humans (or more so), and research attempting to figure out the computational basis of human cognition — that is, how the brain actually carries out its computations. This latter branch of research can be called computational modeling (while the former is often called artificial intelligence or AI).

Almost all contemporary computational modeling research is based on the idea of “neural networks” or, more accurately, “artificial neural networks” (ANNs). An artificial neural network is a simulated network, created using a computer and consisting of nodes and connections between the nodes. The nodes represent or approximate neurons (or groups of neurons), while the connections are analogous to the axons and dendrites connecting neurons. Researchers can test whether their ANNs are successful in carrying out a computation by running simulations on computers. However, to find out whether the ANN is anything like a real network of neurons within the brain requires comparing computational models to findings from real brains. This in turn requires close collaboration or crosstalk between computational modelers and neuroscientists.

Such collaborations have been successful in animal neuroscience research, where neuroscientists have the opportunity to examine working brains up close and in action, for example by implanting electrodes in a living mouse’s brain and observing the patterns of electrical signals as the mouse learns or carries out some tasks [6]. Because it is neither practical nor ethical to perform such experiments with humans, this kind of interdisciplinary (computational and neuroscientific) research on uniquely human types of cognition (such as language) has lagged in comparison to more general types of cognition (such as spatial memory).

Could neuroimaging be the answer to this problem? Neuroimaging provides a safe way to examine the workings of live human brains, but can only produce images on millimeter-scale resolution (for reference, a cubic millimeter could hold at least hundreds of neurons). Furthermore, first generation neuroimaging studies, with their focus on structure-function relationships, were difficult to combine with network models. However, the new methods discussed here take important steps toward bridging that gap. While new structural (DTI) and functional (rs-fcMRI) connectivity methods in MRI cannot reveal relationships between neurons or small groups of neurons as simulated in artificial neural networks, they can be used to investigate relationships between different parts of the brain; these types of relationships can be simulated and examined using a branch of mathematics called graph theory. Furthermore, MVPA provides a sort of “snapshot” of the overall state of a large, complicated network.

These new developments in neuroimaging are bringing cognitive neuroscience closer than ever to compatibility with computational models, and thus bring the field of cognitive science as a whole closer to fulfilling its promise of a full understanding of the mechanisms of human cognition. Who knows what the next generation of neuroimaging methods will bring?

Priya Kalra is a doctoral student at the Harvard Graduate School of Education. She holds a BA in Cognitive Science and an MSc in Functional Neuroimaging.

References:

[1] Klingberg, T, Vaidya CJ, Gabrieli JDE, Moseley ME, and Hedehus M. (1999) Myelination and organization of the frontal white matter in children: A diffusion tensor MRI study. Neuroreport 10(13):2817-2821.

Qiu, D, Tan L-H, Zhou K, et al. (2008) Diffusion tensor imaging of normal white matter maturation from late childhood to young adulthood: Voxel-wise evaluation of mean diffusivity, fractional anisotropy, radial and axial diffusivities, and correlation with reading development. Neuroimage 41(2):223-232.

[2] Wang, L, Goldstein FC, Levey AI, et al.(2011) White matter hyperintensities and changes in white matter integrity in patients with Alzheimer’s disease. Neuroradiology 53(5):373-381.

Patel, S, Mahon K, Wellington R, Zhang J, Chaplin W, and Szeszko PR. (2011). A meta-analysis of diffusion tensor imaging studies of the corpus callosum in schizophrenia. Schizophrenia Research 129(2):149-155.

van Ewijk, H, Heslenfeld DJ, Zwiers MP, Buitelaar JK, and Oosterlaan J. (2012). Diffusion tensor imaging in attention deficit/hyperactivity disorder: A systematic review and meta-analysis. Neuroscience & Biobehavioral Reviews 36(4):1093-1106.

Vandermosten, M, Boets B, Poelmans H, Sunaert S, Wouters J, and Ghesquière P. (2012). A tractography study in dyslexia: Neuroanatomic correlates of orthographic, phonological and speech processing. Brain: A Journal of Neurology 135(3):935-948.

Zhou, Y, Qin L, Chen J, Qian L, Tao J, Fang Y, et al. (2011). Brain microstructural abnormalities revealed by diffusion tensor images in patients with treatment-resistant depression compared with major depressive disorder before treatment. European Journal of Radiology 80(2):450-454.

[3] Keller, TA, and Just MA.(2009) Altering cortical connectivity: Remediation-induced changes in the white matter of poor readers. Neuron 64(5):624-631.

[4] Hoeft, F, McCandliss BD, Black JM, Gantman A, Zakerani N, Hulme C, et al. (2011). Neural systems predicting long-term outcome in dyslexia. Proceedings of the National Academy of Sciences of the United States of America 108(1): 361-366.

Gabrieli, JDE. (2009) Dyslexia: A new synergy between education and cognitive neuroscience. Science 325(5938):280-283.

Nishimoto, S, Vu A, and Gallant J. (2009) Decoding human visual cortical activity evoked by continuous time-varying movies. Journal of Vision 9:667.

Izaki, Y, Takita M, and Nomura M. (2001) Mouse hippocampo-prefrontal paired-pulse facilitation and long-term potentiation in vivo. Neuroreport 12(6):1191-1193.

Links of Interest:

UC Berkeley Scientists “See” Movies in the Mind http://abcnews.go.com/Health/MindMoodNews/scientists-youtube-videos-mind/story?id=14573442#.T96iUlLz-kA

Reading Practice Can Strengthen Brain “Highways” http://www.npr.org/templates/story/story.php?storyId=121253104

What’s Different About the Brains of People with Autism? http://www.npr.org/blogs/health/2012/06/04/154175007/whats-different-about-the-brains-of-people-with-autism

How Your Brain is Like Manhattan http://www.npr.org/blogs/health/2012/03/29/149629657/how-your-brain-is-like-manhattan

Back to Contents

One thought on “Cognitive neuroscience: Connecting neuroimaging and neural nets

Leave a Reply

Your email address will not be published. Required fields are marked *