Keith Bush

A comparison of statistical methods for detecting context-modulated functional connectivity in fMRI

Many cognitive and clinical neuroscience research studies seek to determine how contextual factors modulate cognitive processes. In fMRI, hypotheses about how context modulates distributed patterns of information processing are often tested by comparing functional connectivity between neural regions A and B as a function of task conditions X and Y, which is termed context-modulated functional connectivity (FC). There exist two exploratory statistical approaches to testing context-modulated FC: the beta-series method and psychophysiological interaction (PPI) analysis methods. While these approaches are commonly used, their relative power for detecting context-modulated FC is unknown, especially with respect to real-world experimental parameters (e.g., number of stimulus repetitions, inter-trial-interval, stimulus duration). Here, we use simulations to compare power for detecting context-modulated FC between the standard PPI formulation (sPPI), generalized PPI formulation (gPPI), and beta series methods. Simulation results demonstrate that gPPI and beta series methods are generally more powerful than sPPI. Whether gPPI or beta series methods performed more powerfully depended on experiment parameters: block designs favor the gPPI, whereas the beta series method was more powerful for designs with more trial repetitions and it also retained more power under conditions of hemodynamic response function variability. On a real dataset of adolescent girls, the PPI methods appeared to have greater sensitivity in detecting task-modulated FC when using a block design and the beta series method appeared to have greater sensitivity when using an event-related design with many trial repetitions. Implications of these performance results are discussed.

Modeling reward functions for incomplete state representations via echo state networks

This paper investigates an echo state network (ESN) (Jaeger, 2001 and Maass and Markram, 2002) architecture as the approximation of the Q-function for temporally dependent rewards embedded in a linear dynamical system, the mass-spring-damper (MSD). This problem has been solved utilizing feed-forward neural networks (FNN) when all state information necessary to specify the dynamics is provided as input (Kretchmar, 2000). Time-delayed neural networks (TDNN) solve this problem with finite-size windows of incomplete state information. Our research demonstrates that the ESN architecture represents the Q-function of the MSD system given incomplete state information as well as current feed forward neural networks given either perfect state or a temporally-windowed, incomplete state vector. The remainder of this paper is organized as follows. We introduce basic concepts of reinforcement learning and the echo state network architecture. The MSD system simulation is defined in section IV. Experimental results for learning state quality given incomplete state information are presented in section V. Results for learning estimates of all future state qualities for incomplete state information is presented in section VI. Section VII discusses the potential of the ESN for use in reinforcement learning and provides current and future directions of research.

Robust Reinforcement Learning Control Using Integral Quadratic Constraints for Recurrent Neural Networks

The applicability of machine learning techniques for feedback control systems is limited by a lack of stability guarantees. Robust control theory offers a framework for analyzing the stability of feedback control loops, but for the integral quadratic constraint (IQC) framework used here, all components are required to be represented as linear, time-invariant systems plus uncertainties with, for IQCs used here, bounded gain. In this paper, the stability of a control loop including a recurrent neural network (NN) is analyzed by replacing the nonlinear and time-varying components of the NN with IQCs on their gain. As a result, a range of the NNs weights is found within which stability is guaranteed. An algorithm is demonstrated for training the recurrent NN using reinforcement learning and guaranteeing stability while learning.

Improving the precision of fMRI BOLD signal deconvolution with implications for connectivity analysis.

An important, open problem in neuroimaging analyses is developing analytical methods that ensure precise inferences about neural activity underlying fMRI BOLD signal despite the known presence of confounds. Here, we develop and test a new meta-algorithm for conducting semi-blind (i.e., no knowledge of stimulus timings) deconvolution of the BOLD signal that estimates, via bootstrapping, both the underlying neural events driving BOLD as well as the confidence of these estimates. Our approach includes two improvements over the current best performing deconvolution approach; 1) we optimize the parametric form of the deconvolution feature space; and, 2) we pre-classify neural event estimates into two subgroups, either known or unknown, based on the confidence of the estimates prior to conducting neural event classification. This knows-what-it-knows approach significantly improves neural event classification over the current best performing algorithm, as tested in a detailed computer simulation of highly-confounded fMRI BOLD signal. We then implemented a massively parallelized version of the bootstrapping-based deconvolution algorithm and executed it on a high-performance computer to conduct large scale (i.e., voxelwise) estimation of the neural events for a group of 17 human subjects. We show that by restricting the computation of inter-regional correlation to include only those neural events estimated with high-confidence the method appeared to have higher sensitivity for identifying the default mode network compared to a standard BOLD signal correlation analysis when compared across subjects.

Deconvolution filtering: Temporal smoothing revisited

Inferences made from analysis of BOLD data regarding neural processes are potentially confounded by multiple competing sources: cardiac and respiratory signals, thermal effects, scanner drift, and motion-induced signal intensity changes. To address this problem, we propose deconvolution filtering, a process of systematically deconvolving and reconvolving the BOLD signal via the hemodynamic response function such that the resultant signal is composed of maximally likely neural and neurovascular signals. To test the validity of this approach, we compared the accuracy of BOLD signal variants (i.e., unfiltered, deconvolution filtered, band-pass filtered, and optimized band-pass filtered BOLD signals) in identifying useful properties of highly confounded, simulated BOLD data: (1) reconstructing the true, unconfounded BOLD signal, (2) correlation with the true, unconfounded BOLD signal, and (3) reconstructing the true functional connectivity of a three-node neural system. We also tested this approach by detecting task activation in BOLD data recorded from healthy adolescent girls (control) during an emotion processing task. Results for the estimation of functional connectivity of simulated BOLD data demonstrated that analysis (via standard estimation methods) using deconvolution filtered BOLD data achieved superior performance to analysis performed using unfiltered BOLD data and was statistically similar to well-tuned band-pass filtered BOLD data. Contrary to band-pass filtering, however, deconvolution filtering is built upon physiological arguments and has the potential, at low TR, to match the performance of an optimal band-pass filter. The results from task estimation on real BOLD data suggest that deconvolution filtering provides superior or equivalent detection of task activations relative to comparable analyses on unfiltered signals and also provides decreased variance over the estimate. In turn, these results suggest that standard preprocessing of the BOLD signal ignores significant sources of noise that can be effectively removed without damaging the underlying signal.

Subthreshold-Seeking Behavior and Robust Local Search

Subthreshold-seeking behavior occurs when the majority of the points that an algorithm samples have an evaluation less than some target threshold. We characterize sets of functions where subthreshold-seeking behavior is possible. Analysis shows that subthreshold-seeking behavior, when possible, can be increased when higher bit precision is used with a bit climber search algorithm and a Gray code representation. However, higher precision also can reduce exploration. A simple modification to a bit-climber can improve its subthreshold-seeking behavior. Experiments show that this modification results in both improved search efficiency and effectiveness on common benchmark problems.

Distributed Neural Processing Predictors of Multi-dimensional Properties of Affect

Recent evidence suggests that emotions have a distributed neural representation, which has significant implications for our understanding of the mechanisms underlying emotion regulation and dysregulation as well as the potential targets available for neuromodulation-based emotion therapeutics. This work adds to this evidence by testing the distribution of neural representations underlying the affective dimensions of valence and arousal using representational models that vary in both the degree and the nature of their distribution. We used multi-voxel pattern classification (MVPC) to identify whole-brain patterns of functional magnetic resonance imaging (fMRI)-derived neural activations that reliably predicted dimensional properties of affect (valence and arousal) for visual stimuli viewed by a normative sample (n = 32) of demographically diverse, healthy adults. Inter-subject leave-one-out cross-validation showed whole-brain MVPC significantly predicted (p < 0.001) binarized normative ratings of valence (positive vs. negative, 59% accuracy) and arousal (high vs. low, 56% accuracy). We also conducted group-level univariate general linear modeling (GLM) analyses to identify brain regions whose response significantly differed for the contrasts of positive versus negative valence or high versus low arousal. Multivoxel pattern classifiers using voxels drawn from all identified regions of interest (all-ROIs) exhibited mixed performance; arousal was predicted significantly better than chance but worse than the whole-brain classifier, whereas valence was not predicted significantly better than chance. Multivoxel classifiers derived using individual ROIs generally performed no better than chance. Although performance of the all-ROI classifier improved with larger ROIs (generated by relaxing the clustering threshold), performance was still poorer than the whole-brain classifier. These findings support a highly distributed model of neural processing for the affective dimensions of valence and arousal. Finally, joint error analyses of the MVPC hyperplanes encoding valence and arousal identified regions within the dimensional affect space where multivoxel classifiers exhibited the greatest difficulty encoding brain states – specifically, stimuli of moderate arousal and high or low valence. In conclusion, we highlight new directions for characterizing affective processing for mechanistic and therapeutic applications in affective neuroscience.

Decoding the Traumatic Memory among Women with PTSD: Implications for Neurocircuitry Models of PTSD and Real-Time fMRI Neurofeedback

Posttraumatic Stress Disorder (PTSD) is characterized by intrusive recall of the traumatic memory. While numerous studies have investigated the neural processing mechanisms engaged during trauma memory recall in PTSD, these analyses have only focused on group-level contrasts that reveal little about the predictive validity of the identified brain regions. By contrast, a multivariate pattern analysis (MVPA) approach towards identifying the neural mechanisms engaged during trauma memory recall would entail testing whether a multivariate set of brain regions is reliably predictive of (i.e., discriminates) whether an individual is engaging in trauma or non-trauma memory recall. Here, we use a MVPA approach to test 1) whether trauma memory vs neutral memory recall can be predicted reliably using a multivariate set of brain regions among women with PTSD related to assaultive violence exposure (N=16), 2) the methodological parameters (e.g., spatial smoothing, number of memory recall repetitions, etc.) that optimize classification accuracy and reproducibility of the feature weight spatial maps, and 3) the correspondence between brain regions that discriminate trauma memory recall and the brain regions predicted by neurocircuitry models of PTSD. Cross-validation classification accuracy was significantly above chance for all methodological permutations tested; mean accuracy across participants was 76% for the methodological parameters selected as optimal for both efficiency and accuracy. Classification accuracy was significantly better for a voxel-wise approach relative to voxels within restricted regions-of-interest (ROIs); classification accuracy did not differ when using PTSD-related ROIs compared to randomly generated ROIs. ROI-based analyses suggested the reliable involvement of the left hippocampus in discriminating memory recall across participants and that the contribution of the left amygdala to the decision function was dependent upon PTSD symptom severity. These results have methodological implications for real-time fMRI neurofeedback of the trauma memory in PTSD and conceptual implications for neurocircuitry models of PTSD that attempt to explain core neural processing mechanisms mediating PTSD.

Evidence-based modeling of network discharge dynamics during periodic pacing to control epileptiform activity

Deep brain stimulation (DBS) is a promising therapeutic approach for epilepsy treatment. Recently, research has focused on the implementation of stimulation protocols that would adapt to the patients need (adaptive stimulation) and deliver electrical stimuli only when it is most useful. A formal mathematical description of the effects of electrical stimulation on neuronal networks is a prerequisite for the development of adaptive DBS algorithms. Using tools from non-linear dynamic analysis, we describe an evidence-based, mathematical modeling approach that (1) accurately simulates epileptiform activity at time-scales of single and multiple ictal discharges, (2) simulates modulation of neural dynamics during epileptiform activity in response to fixed, low-frequency electrical stimulation, (3) defines a mapping from real-world observations to model state, and (4) defines a mapping from model state to real-world observations. We validate the real-world utility of the models properties by statistical comparison between the number, duration, and interval of ictal-like discharges observed in vitro and those simulated in silica under conditions of repeated stimuli at fixed-frequency. These validation results confirm that the evidence-based modeling approach captures robust, informative features of neural network dynamics of in vitro epileptiform activity under periodic pacing and support its use for further implementation of adaptive DBS protocols for epilepsy treatment.

Brain States that Encode Perceived Emotion are Reproducible but their Classification Accuracy is Stimulus-Dependent

The brain state hypothesis of image-induced affect processing, which posits that a one-to-one mapping exists between each image stimulus and its induced functional magnetic resonance imaging (fMRI)-derived neural activation pattern (i.e., brain state), has recently received support from several multivariate pattern analysis (MVPA) studies. Critically, however, classification accuracy differences across these studies, which largely share experimental designs and analyses, suggest that there exist one or more unaccounted sources of variance within MVPA studies of affect processing. To explore this possibility, we directly demonstrated strong inter-study correlations between image-induced affective brain states acquired 4 years apart on the same MRI scanner using near-identical methodology with studies differing only by the specific image stimuli and subjects. We subsequently developed a plausible explanation for inter-study differences in affective valence and arousal classification accuracies based on the spatial distribution of the perceived affective properties of the stimuli. Controlling for this distribution improved valence classification accuracy from 56% to 85% and arousal classification accuracy from 61% to 78%, which mirrored the full range of classification accuracy across studies within the existing literature. Finally, we validated the predictive fidelity of our image-related brain states according to an independent measurement, autonomic arousal, captured via skin conductance response (SCR). Brain states significantly but weakly (r = 0.08) predicted the SCRs that accompanied individual image stimulations. More importantly, the effect size of brain state predictions of SCR increased more than threefold (r = 0.25) when the stimulus set was restricted to those images having group-level significantly classifiable arousal properties.