MODELING AND THEORY in NEUROIMAGING

In my talk I will outline recent work in our laboratory that has examined the nature of risk and reward signals and their relation to learning. We first examine the way in which value is represented in the context of decision making, using a simple mixed-gamble paradigm. This work shows that aversion to losses in the context of decisions is not driven by engagement of negative emotional systems, but instead by decreased activity in regions coding for reward. Further, individual differences in behavioral loss aversion are strongly predicted by differences in 'neural loss aversion'. A second study examines how tradeoffs between increasing risk and potential reward are computed, using the Balloon Analog Risk Task. This work shows that both reward-sensitive regions and regions involved in cognitive control show increased engagement as risk and potential reward increase. Amygdala is sensitive to experienced losses, though its response habituates with experience. Finally, I will outline work suggesting that the same areas engaged by monetary rewards are also sensitive to informational feedback, suggesting that these systems respond to a wide range of behaviorally-relevant outcomes. This work was recently reported in SCIENCE, The Neural Basis of Loss Aversion in Decision-Making Under Risk Tom et al. Science 26 January 2007: 515-518

The Perceptual Expertise Netwok (PEN) is a group of cognitive neuroscientists working collaboratively since 2001 to explore how "different" brains approach object recognition and categorization. I will report on our efforts to study the ability of individuals with prosopagnosia to acquire expertise with novel objects. Three training case studies suggest as much heterogeneity in prosopagnosics' ability to learn about novel objects ("Greebles") as there is in their face and object recognition difficulties. However, one commonality across all three cases is that, regardless of whether or not they can reach our criterion for Greeble expertise, they do not seem to recruit holistic processing. I will also report on a study of CK, an individual with "reverse prosopagnosia": we took advantage of this very rare deficit to address the difficult question of what counts as a "face-like" stimulus to the face recognition module, assuming there is such an entity. Together these studies represent a new step towards controlled studies of learning in visual agnosia, taking us beyond conjectures and anecdotal evidence.

Functional MRI has revealed several cortical regions in the ventral visual pathway in humans that exhibit a striking degree of functional specificity: the fusiform face area (FFA), parahippocampal place area (PPA), and extrastriate body area (EBA). I will briefly review this work and then discuss more recent studies that investigate the specificity, origins, and generality of domain specificity in the cortex. In particular these studies ask i) how specialized is the FFA for faces and what exactly does it do with faces?, ii) how do domain specific regions arise in the cortex and is extensive experience ever sufficient to create them?, and iii) are domain specific regions of cortex found only in the visual system, or can they sometimes be found for very abstract high-level cognitive functions as well?

Independent component analysis (ICA) has already found many applications in basic signal analysis. I will demonstrate that ICA is also useful in a wide range of higher-level tasks, such as text analysis, multimedia modeling, etc. There are strong indications that the brain implements ICA-like representations at the perceptual level-should we look for such representations at the cognitive level as well? The talk reviews our ICA results in chat room analysis, webmining, network decomposition, and in neuroinformatics.

One critical and highly adaptive aspect of the human visual system is our seemingly effortless ability to identify objects even when only partial and often very sparse visual information is presented to the observer. The neural processes responsible for filling-in of missing information that enable eventual object-recognition under partial viewing conditions (e.g. fog, occlusion, camouflage, poor lighting) have come to be referred to as 'Perceptual Closure' processes. In a series of studies, we have used the excellent temporal resolution of high-density electrical recordings in combination with the precise spatial localization afforded by functional imaging to investigate the spatio-temporal dynamics of these perceptual closure processes. We have defined the neural circuitry responsible for 'closure' computations and roughly defined the timecourse of interactions across this circuit. In a series of related studies, we have investigated the neural circuitry involved in the analysis of illusory contour stimuli to assess whether these processes are the same as those involved in perceptual closure, as is often claimed. We find that similar circuits are active but with very different timecourses. From these results, a two-stage model of object-processing has been devised and will be discussed.

As we listen casually to sentences during conversation or while we watch TV, our brains carry out highly complex computations on the incoming signal. Central among these is high-speed syntactic analysis which we dorather effortlessly. Linguists study the nature of these computations; neurolinguists investigate their neural implementation. How languagemechanisms might be organized in the brain, what counts as relevant evidence, and how it is adduced, will be the topic of this Talk. It will be about what we know about the organization of syntax in the brain, and how we came to know it.
I will start with a quick tutorial for the uninitiated, and present some theoretical considerations that motivate central syntactic principles. I will then present evidence regarding their neurological instantiation. First among these will be syntactic movement (a k a grammatical transformations). I will discuss neurological studies of this rule system in healthy subjects, carried out through fMRI tests of receptive syntax, and show how different applications of movement (e.g., questions, topicalization) activate similar cerebral loci (mostly in the left inferior frontal gyrus [Broca's region], and in both temporal lobes [Wernicke's region]).
Next, these results will be juxtaposed to those coming from behavioral and anatomical investigations of Broca's aphasic patients, usually subsequent to left-hemispheric stroke. When looked at superficially, the picture that arises is not pretty: it presents seemingly unruly, sometimes strange, patterns; further puzzles appear when cross-linguistic differences are thrown into the pot (e.g., aberrant performances of aphasic speakers of German/Dutch vs. Spanish, or Chinesevs. English). I will propose solutions to some of these puzzles, and try to show that the disarray in the empirical record reduces drastically when the right methodological considerations (statistical as well as linguistic) are introduced. I will then argue for a unified account, that suggests high regularity in impairment patterns in aphasia, and that also extends to the previously presented fMRI data.
Finally, I will present results from very recent fMRI studies in healthyadults, of syntactic contrasts that activate areas other than the traditional language areas. Of particular are intra-sentential dependency relations other than syntactic movement, that make demands on Working Memory, yet are surprisingly located in the frontal lobe of the right cerebral hemisphere. I will use these data to consider a broader set of questions how syntax is organized in the brain, whether or not it is distinct from other cognitive systems, and what the best method to localize it is. Bringing in certain brain mapping principles established by students of vision, as well as some cytoarchitectonic considerations, I will try to imagine how a brain map for syntax might be constructed.

Functional MRI has experienced explosive growth in the past 12 years because it is easy to use and it produces results that are robust and interpretable. The most straightforward fMRI methodology involves collection of a time series of images with at least two interleaved conditions lasting from less than 2 sec (event-related design) to more than 10 sec (blocked design). A statistical comparison is then made on a voxel-wise basis between the magnitude signal intensity changes induced by the conditions. While this methodology is extremely successful, I believe that it does not fully tap into the potential information contained in the fMRI time series data. Much of the research in my lab is aimed at developing methods by which the questions being asked, the paradigms, the imaging methods, and the processing techniques are tailored to increase the temporal resolution, spatial resolution, and interpretability of fMRI. The additional information that can be extracted include: relative neuronal timing information on the order of 100 ms, baseline spontaneous neuronal information, true volumes of activation, and magnitude information that is correlated with the degree of neuronal activity.
I will discuss techniques for calibrating the signal change magnitudes and latencies such that the spatial and variability resulting from non-neuronal factors (i.e. the variation in the vessel architecture) are reduced or eliminated. Related to calibration method development is work aimed at more fully characterizing the relationship between neuronal activity and the measured MRI signal change. Recent research in my lab aimed at characterizing the nonlinearites in the fMRI response will be demonstrated. In addition, it is generally accepted that baseline data has information about neuronal activity. I will discuss some of our recent work aimed at better understanding the origins of these fluctuations as well as more robustly extracting neuronal information from baseline signal. Lastly, I will discuss results and the potential for measuring neuronal activity directly using MRI: known as neuronal current imaging or magnetic source imaging.

Using simulations I will demonstrate the improved signal-detection performance of multivariate over univariate data analysis approaches in neuroimaging, and argue for a central role for data-driven discovery in functional neuroimaging based on flexible, adaptive, multivariate models (e.g., Canonical Variates Analysis (CVA), Support Vector Machines(SVM), etc.). I will describe the NPAIRS resampling framework that we recently introduced (Strother et al., Neuroimage 15:747, 2002; Kjems et al., NI 15:772, 2002) that uses a combination of cross-validation and delete-d jackknife resampling to obtain prediction and reproducibility metrics with a training set that acts as an exploratory modeling phase followed by a confirmatory phase using an independent test set. We have presented preliminary data (LaConte et al., NI 18:10, 2003; Shaw et al., NI 19:988, 2003) that these metrics provide a data-driven alternative to simulations and Receiver Operating Characteristic curves for evaluating signal detection performance. I will demonstrate the use of these NPAIRS metrics with two fMRI data sets: high spatial resolution, single-slice, event-related studies of the cat primary visual cortex at 4.7T in collaboration with Dr. Seong-Gi Kim, and whole-brain, block-design studies of parametric static force between the thumb and forefinger at 1.5T in collaboration with Drs. Xiaoping Hu and James Ashe. I will end with some thoughts on consensus modeling (Hansen et al., NI 13:1212, 2001) as an alternative to choosing a single 'optimal' preprocessing and data analysis modeling approach in functional neuroimaging.