2007 IGERT Student Research Symposium Abstracts

Synchronous neuronal firing can serve to amplify signals to downstream brain regions and may play a role in the neural representation of sensory information. The phenomenon of stochastic synchrony, whereby correlated noisy inputs induce synchronous firing in otherwise uncoupled cells, is ubiquitous but poorly understood. The mathematical theory of weakly coupled oscillators predicts that neurons displaying Hodgkin's class 2 behavior, such as cortical interneurons and olfactory mitral cells (MCs), will synchronize more readily in the presence of correlated noise than those showing Hodgkin's class 1 behavior, such as cortical pyramidal cells and olfactory granule cells. Moreover, experimental results in mouse olfactory bulb show that blockade of low threshold A-type potassium currents in MCs can effectively switch the cell's excitability profile from class 2 to class 1. This presentation will provide a concise and easily digestible introduction to the underlying mathematics of phase oscillators, the role of intrinsic currents in oscillator dynamics, and applications to sensory processing in the olfactory bulb.

A fundamental challenge that neural systems must overcome is the task of discriminating between similar but distinct stimuli. The olfactory system presents a unique challenge because odor presentation produces a distributed pattern of activity such that the spatial location of neurons is not well correlated with their odor response profile. While the primary output neurons of the bulb receive extensive reciprocal, dendrodendritic inhibitory connections, traditional mechanisms of contrast enhancement utilizing the spatial location of neurons as a correlate of their response profiles (such as in the retina) are not effective in enhancing contrast in this system. Our experimental results provide evidence to support a novel method for specifying lateral inhibition such that the strength of inhibition is determined by the activity of both the pre- and postsynaptic neurons. In this way, inhibitory output of the network can be dynamically remapped based on the pattern of activity within the network. Computational modeling results demonstrate that this algorithm is very effective in enhancing contrast even in the absence of any spatial structure within the network. We extended this algorithm to image processing where it was also effective in enhancing contrast, even if spatial structure was removed by randomizing pixels before processing and then unrandomizing after processing.

Neural probes do not interface chronically with the neurons in the brain due to biocompatibility issues. These problems arise at the tissue-electrode interface such as the reduction of neuronal density around the probe and the glial sheath encapsulation.

The aim of this research is to modify the surface of these probes with the neural attractive L1 protein with the goal of improving biocompatibility. We hypothesize that maintaining neuron-electrode proximity is instrumental for obtaining long-term stable recordings. To test this hypothesis, in vivo work is being performed in the barrel cortex of rat models to demonstrate the direct relationship between recording and neuronal density around theimplant.

The messages a neuron delivers to its postsynaptic targets are determined by the coupling between electrical activity and neurotransmitter release. Most neurons signal digitally -- action potentials are required to evoke release of transmitter. In non-spiking neurons, subthreshold changes in membrane potential produce graded changes in transmitter release. Here we show that mitral cells of the accessory olfactory bulb release can glutamate from their dendrites in both graded and action potential-dependent fashion. Moreover, the dominant mode of transmitter release can be altered by either pharmacological or endogenous activation group I metabotropic glutamate receptors These results provide the first demonstration of a CNS neuron capable of both digital and pure analog transmitter release, and represent a new form of synaptic modification in which graded transmission -- and hence the capacity of neurons to convey analog signals -- is switched on or off by changes in activity.

Reinforcement-based algorithms for policy learning generally rely on an error signal both to indicate the delivery of unpredicted rewards and to strengthen the actions that preceded those rewards. As highlighted in a recent discussion (Niv et al., Nature Neuroscience, 2006) the calculation of this error depends on the particular learning algorithm employed; electrophysiological evidence in primates (Morris et al., Nature Neuroscience, 2006) suggests that dopaminergic firing may most resemble the action-specific prediction errors employed in "state-action-reward-state-action" (SARSA) methods, as opposed to the exclusively state-based errors found in actor-critic models. To illustrate this result, we describe the behavior of both actor-critic and SARSA-based models in a training situation similar to the Morris et al. experiment.

In the zebrafish, the central pattern generator for swimming is composed ofinterneurons in which the resistance varies continuously with their position in the spinal cord. We examined how this heterogeneity may affect swimming at different frequencies using a network model consisting of bursting Izhikevich neurons. We found that resistance heterogeneity could be important in producing a larger range of swimming strengths at slow swimming speeds. We also performed a sensitivity analysis to ascertain the dependence of the central pattern generator output to the distribution ofresistances and other parameters affecting the dynamics.

We undertake the study of Electrical Deep Brain Stimulation (EDBS), a treatment for the symptoms of Parkinson's Disease, from the perspective of control theory. Recent advances in microelectrode array technology have opened the possibility of utilizing many independent electrodes, not only for stimulus but also for feedback. The goal of EDBS is to correct the pathological synchronization of neural signals. We put this objective into a control theory context and develop feedback-based control algorithms that desynchronize simple models of neural populations using control waveforms, proportional to the degree of synchrony, that minimize the total delivered energy.

How are shapes and objects in the environment represented and recognized? The research on the constituent part structure of an object enables us to make certain predictions, for example, how movement of one part may affect the percept of the entire shape. For instance, object's part boundaries are indicated by negative minima of curvature, shown in both static (Hoffman & Richards, 1984) and moving, part-articulating (Barenholtz & Feldman, 2006) shapes. Furthermore, while some shapes move in a rigid manner, such as when a book falls on the floor, others move non-rigidly, transforming their shapes in the process, such as when a long-limbed kitten bends its legs at the knees and hips while walking, or decides to curl its tail.

I am investigating shape representation of non-rigidly moving objects with part structure using an apparent motion technique. (Apparent motion may be elicited by alternately showing two static, often identical, stimuli at spatially close locations in the field of vision. There is tendency to perceive a single object, as opposed to multiple objects spontaneously appearing and disappearing in space). Overarching questions include: What are the circumstances under which a smooth transformation of a non-rigidly moving object is perceived? How does part structure affect the percept of smooth shape transformation? More generally, what does this tell us about the underlying representation of visual shape? Addressing these questions in the context of human perception is of relevance to the study of computer vision, for example in interpolating a smooth motion trajectory between two consecutive static frames of a video sequence. This project is inspired by the idea that moving biological entities engage shape representation principles in the brain of a human observer that are distinct from those used in interpreting rigidly moving inanimate objects, and may reveal a more complex system of shape representation.

Neural responses to oriented stimuli in primary visual cortex (V1) are organized into columns of neurons that respond best to stimuli presented at a preferred angle to a given eye. The columnar organization of V1 is that of coherent maps of orientation and eye preference. We characterize the proteomic differences across these maps using differential gel electrophoresis (DIGE) on functionally defined column types. To reduce variation in this analysis, we are applying image analysis tools to map structures. Characterizing the local geometry of functional maps in V1 will allow us to improve our proteomic analysis through standardization of tissue sampling.

We present a biologically inspired neural network model capable of successfully performing the antisaccade task, a prominent experimental paradigm used as an index of response inhibition. The model employs a threshold activation function and incorporates pools of simple processing units whose output activations over time approximate electrophysiological recordings in saccade and fixation neurons in the superior colliculi and the frontal eye fields (Munoz and Everling, 2004). Importantly, the model replicates a number of empirically observed behavioral measures including latencies for correct antisaccades and erroneous prosaccades, as well as prosaccade error rates. Simulation results will be presented and discussed.

Neuromodulation is one way for a small network to produce a variety of behaviors. Recent work has shown that networks with different underlying properties can produce similar output. This raises the question: are the effects of neuromodulators consistent if the underlying properties of the network are not? To examine this question we use the dynamic clamp to construct two-cell networks from crab neurons. We then study the effects of varying the synaptic and cellular parameters on the networks in control conditions and in serotonin.

Dramatic refinement of tonotopic maps occurs before hearing onset. We posit that the spontaneous bursts of activity within the auditory system drive this refinement through spike timing dependent plasticity (STDP). To test this proposition, we designed an STDP model that applies nonlinear Temporally Asymmetric Hebbian learning rules to update synaptic weights of inputs with differing delay lines. We explored the sensitivity of this system to different inputs by using input patterns that included permutations with regards to the type (i.e. homogeneous or inhomogeneous Poisson processes) and independence of bursting patterns and spike patterns within those bursts.

We investigate existence and stability of traveling pulses in a one-dimensional neural network with recurrent excitation. A network model uses non-local integro-differential equations whose integral kernel represents spatial distribution of synaptic weights. The latter is decomposed into an exponential representing local connections and a set of delta functions representing long-range patchy connections. Solving for pulse solutions of the system, we determine wave-speed's relation to threshold. Stability is determined by zeroes of the associated Evans function. We begin with the homogeneous case ($e = 0$), and then extend this using perturbation theory to the inhomogeneous case, where long range connections are periodically anisotropic. As the amplitude of anisotropy in long range connections increases, eventually, no stable pulses exist.

Learning from experience always involves striking a tradeoff between bias and variance. High bias models can learn from little data, but may never learn accurate predictions. High variance models can become accurate, but may require unrealistic amounts of data to do so. In this work, I study the problem of learning accurate motion models efficiently in a robotic task. Specifically, the model strikes a bias-variance balance by generalizing experience across parts of the state space with similar surface appearance. Classical Markov decision process (MDP) models of environments have high variance because they do not support the critical idea of generalization between states. Our technique assumes that states that appear similar to the visual system will have similar transition dynamics and uses this assumption to increase bias. We implemented this approach in both simulated and robot domains and found a significant decrease in the learning time needed to achieve near-optimal task performance.

The deep tendon reflex is a fundamental aspect of a neurological examination. The device for quantifying deep tendon reflex is a MEMS based device that accurately quantifies both reflex response and latency. The device has relevancy to central and peripheral nervous system traumas. The MEMS accelerometers are attached to a set anchor point near the ankle and reflex hammer swing arm. The reflex amplitude is based on the maximum acceleration of the reflex response. The temporal disparity between hammer strike and response defines the reflex latency. Quantified data collected from MEMS accelerometers is transmitted by a portable computer.

Virtual Proprioception

Virtual proprioception represents a novel means of developing cortical reorganization of alternative strategies for hemiparetic gait. Fundamentals of the device are motor control plasticity and the Mussa-Ivaldi aftereffect. Two 3D MEMS accelerometers are placed on both the femur (upper leg) of the ipsilesional and contralateral limb. The acceleration data from the two 3D MEMS accelerometers is feedback to the user in real time by visual output from a head mounted display after processing from a wearable computer. Given the virtual proprioception feedback the user can then adjust the original gait while walking to an improved alternative gait strategy.

Models of verbal working memory assume modality-specific representations are translated into phonological representations upon entering into a working memory system that supports maintenance and rehearsal. Once encoded phonologically, the processing is assumed to be amodal. We consider a 1-back memory paradigm to focus on the phonological storage component of verbal working memory. We consider passive listening versus passive viewing of words, 1-back memory of heard words versus passive listening, 1-back memory of seen words versus passive viewing, and finally, 1-back memory of heard words minus passive listening, versus 1-back memory of seen words minus passive viewing. Our results show that that there are modest but significant differences in 1-back memory for heard versus seen words. Such differences were found bilaterally in temporal cortex, supplementary motor cortex and the anterior cingulate, suggesting differences in phonological maintenance between modalities. We also found significant sustained effects bilaterally in temporal cortex. Overall our results demonstrate that brief maintenance processing of single-words is modulated by input modality.

Patterned neuronal activity and molecular cues interact to varying degrees to shape the anatomy of the developing brain. In the visual system, a long standing hypothesis has been that patterns of retinal activity are essential for the development of ocular dominance columns. We confirm that in addition to forming early in development, well before sensory experience, ocular dominance columns develop independently of the spontaneous retinal activity which acts as a surrogate for visual experience before it begins. To supplement our anatomical data, we have used in vitro calcium imaging to explore the maturation of functional architecture at the column boundary. Our results recast a classic model of development, that early crude anatomical connectivity is refined through sensory experience. Instead, we propose that development is more precise, depending on molecular markers to guide anatomical projections to their destinations.

Since Leonardo da Vinci, artists and psychologists have known that, all other things being equal, brighter stimuli are perceived to be closer than darker stimuli. Using a laser range scanner, we have shown that this perceptual bias is adaptive: in natural scenes, brighter stimuli are in fact more likely to be closer. We present evidence that this statistical tendency is caused by shadows in complex natural scenes. We then show the potential that this cue has in computer vision applications. Finally, we present single-cell recording data from awake behaving macaques that shows that this statistical trend is exploited in the visual system as early as V1. Specifically, cells that prefer near disparities tend to prefer brighter stimuli.

In cases of spinal cord injury that lead to paralysis, there is a functional disconnection between healthy brainstem and healthy spinal cord tissue. The crux of the problem is that the descending axons that normally convey motor commands from brain to spinal cord are no longer operable. Two projects are described exploring the use of nanotechnology to investigate molecular mechanisms underlying regeneration and as a platform to evaluate therapeutic and pharmacological interventions: (1) exploring the use of DNA- functionalized nanoparticles to deliver neural repair genes to brainstem neurons damaged in spinal cord injury. (2) evaluation of nanowire arrays to serve as a brain-machine interface to elicit and control locomotor movements.

This work aims to establish the operating parameters and nanowire array dimensions required for the effective recording and stimulation of neural activity. The conjunction of these two approaches offers synergistic opportunities that are applicable not only to spinal cord injury, but also to a broader variety of neurological and neurodegenerative disorders.

The provision of wheeled mobility and seating interventions can be complex especially when considering people with complex needs, environmental factors, and wide array of product interventions. The availability of qualified practitioners with specialty expertise in this area is limited especially outside urban areas therefore people are potentially isolated from rehabilitation services due to geography or physical limitations. Distance can mean excessive travel time and increased burdens. Telerehabilitation (TR) has been described and discussed as a method to potentially alleviate some of these issues. A research study has been launched at the University of Pittsburgh to more objectively determine the effectiveness and accuracy of using a TR consultation model for procuring appropriate wheeled mobility and seating devices for individuals with mobility impairments. Anticipated outcomes for this study are similar magnitudes if improved function as measured by the Functioning Everyday with a Wheelchair (FEW) for two independent samples; people receiving an in-person assessment by an expert clinician and people receiving an in-person assessment by a generalist clinician with expert consultation via TR.

Our ability to adapt our movements based on visual information demonstrates that the visuomotor system is highly adept at learning. Most current theories of motor learning postulate that the amount of learning that occurs is proportional to the amount of induced error. However, depending on the statistics of the environment it appears that people can adopt different learning strategies, using both proportional learning strategies, as well as categorical learning strategies. Investigation of trial-by-trial learning during tasks including changes in visual feedback are few; by using this technique, we can evaluate the behavior of visuomotor learning strategies using different environmental statistics.

Object learning and recognition in infants are interactive processes that simultaneously construct, apply, and refine models of visual phenomena. As part of an effort to develop a computer vision system with similar characteristics, I'll present two component techniques designed to support this behavior. The first learns object models as dynamically-connected views. The second is a clustering algorithm that uses the learned models to interpret images; its design aims to allow synthesis of bottom-up and top-down cues. Both techniques make of Dirichlet Process-based models, since these techniques feature flexibility that supports our overall goal. Additionally, the second algorithm is notable for its novel approach to applying these techniques to graphs.

The concept of reliability is defined traditionally by philosophers of science as a feature that an experiment has when its repeated use results in data that can be used to arrive at true claims about phenomena. In contrast, the concept of validity is taken to correspond roughly to that of generalizability -- a feature that a scientific claim has when it is based on laboratory data but can be used to explain features of phenomena external to the laboratory. I demonstrate how reliability and validity are two competing goals of the experimental process in the context of neurobiological experiments on learning. I suggest that the express commitment to reliability in this research area has resulted in claims about organism-level learning that have limited application beyond the laboratory. The positive component of talk consists of specific proposals that I offer as guidelines for resolving the competition between reliability and validity in the context of experimental design.

Neural circuits must be stable in some respects yet plastic in others. They must maintain appropriate output in the face of ion channel and receptor turnover, but be plastic enough so that the animal can respond to environmental change. For cold-blooded animals (e.g. crabs, worms, flies), the external temperature changes the kinetics of processes throughout the nervous system, and the nervous systems of these animals must maintain proper function in spite of these changes. We studied the pyloric rhythm of the crab, Cancer borealis, stomatogastric nervous system (STNS) as a function of temperature. We divided a number of animals into three groups, and acclimated them to different temperatures (7, 11, or 19^o C) for at least two weeks. For each animal, we then acutely varied the temperature of the isolated STNS in vitro from 5 to 31^o C. Within a given animal, network frequency was approximately a linear function of acute temperature in the range 7 to 23^o C. For all three temperature acclimation groups (7^o C,11^o C, and 19^oC), their mean network frequencies at T = 7^o,11^o, 15^o, 19^o, and 23^o C, were approximately 0.7, 1.2, 1.8, 2.4, and 2.9 +/- 0.1 Hz, respectively. Within this temperature range, the frequency vs. temperature relationship was remarkably similar/stable both within animals of the same acclimation group and across animals of all three acclimation groups (7^o C, 11^o C, and 19^o C) (there were no statistically significant differences in cycle frequency between the 7^o-,11^o-, and 19^o-acclimated animals in the 7 to 23^o C range). However, the 19^o-acclimated animals were more likely to burst robustly at 31^o C than the 7^o-acclimated animals (p<0.05, Fisher's exact test with Bonferroni correction). In addition, the 19^o-acclimated animals could cycle at maximum network frequencies of 4.7 +/- 0.1 Hz (at 31^oC), which were significantly faster than the maximum network frequencies of the 7^o- and 11^o-acclimated animals, which were 3.0 +/- 0.1 Hz and 2.9 +/- 0.2 Hz respectively (p<0.005, t-test, with Bonferroni correction). These data suggest that long-term temperature perturbations can induce alterations that extend the range of network performance. Importantly, these compensations occur without disrupting the ability of the network to produce stable output in its normal operating regime.

The goal of the RUBI project is to develop a social robot (RUBI) that can interact with children and teach them in an autonomous manner. We are currently focusing on the problem of teaching 18-24 month old children. RUBI teaches the children by playing question/answer formatted games with them on a touch screen on her stomach. After each game, RUBI has to decide which game to switch to next. We are using Reinforcement Learning Methods to have her learn which games the children prefer and play them more often. Simulation results have already shown that using very simple techniques such as epsilon-greedy Monte Carlo to choose the next game to play could be very effective in maximizing the children's attention and continuing interaction with the games.

We present a computer audition system that can both annotate novel audio tracks with semantically meaningful words and use a semantic query to retrieve relevant tracks from database of unlabeled audio content. We consider the related tasks of content-based audio annotation and retrieval as one supervised multi-class problem in which we model the joint probability of acoustic features and words.

Recent studies have demonstrated that the visual observation of another individual's trial-and-error adaptation to unknown force perturbations can provide a forward prediction of environment dynamics. Previously, researchers have had subjects first observe another naive actor's reach adaptation in an unknown force field environment, then execute the same reaching task in either the same or a different force field environment. A comparison of errors averaged over several movements during the execution phase revealed either an adaptive or interfering effect, depending upon the imposed forces. The mechanism by which visually observed information is transformed into the error signal driving this adaptive effect is currently unknown. We propose a novel trial-by-trial experimental and analytical approach, using interspersed observation, execution, and error-clamped trials, to characterize and model the incremental effects of motor observation upon predictive motor control of horizontal-plane reaching movements.

Attention is crucial to natural vision, contributing to perceptual experience and guiding saccadic eye movements. A central issue for computational models of natural vision is how limited attentional resources are distributed to handle multiple task requirements. We studied active tasks where subjects counted, pointed to, or simply looked at patterns of dots, while a probe (oriented Gabor) was flashed randomly. Identification of high-contrast Gabors was impaired during counting. Pointing and looking had modest effects. These results show a strong bottleneck limiting what can be identified during fixation pauses. Detecting, selecting, and the localization of objects is spared.

Most theories of speech perception assume phoneme identification as a critical step in lexical identification. Several neuroimaging studies indicate that posterior aspects of the left superior temporal gyrus and sulcus, as well as left middle temporal gyrus are involved in phonemic processing. However, the capacities of the right homologues of these regions to support phonemic processing are not well understood. The extent to which the right hemisphere may or may not be able to organize gradient acoustic stimuli into phonemic categories is discussed, paying particular attention to evidence from neuropsychology.

The human brain is constantly bombarded with sensory information originating from disparate sources, challenging the nervous system with the task of discerning what signals came from the same source and therefore should be integrated. We report a novel divided attention discrimination task that required subjects to report the number of simultaneously presented auditory, visual, and tactile pulsations. This work shows that human auditory-visual-tactile perception is consistent with a Bayesian ideal observer, indicating that the rule used by the nervous system for when and how to combine auditory, visual and tactile signals is statistically optimal.

We examined visuospatial transformations of rotated stimuli depicting either animals or inanimate objects. Two behavioral tasks were used. One involved judgments of stimulus similarity, while the other involved judgments of chirality, or handedness. For the latter task, we predicted that perspective-taking would be the predominant strategy when the stimuli depicted animals, but not objects. Notably, reaction time patterns differ from results obtained in other studies using depictions of human bodies. Our results suggest that the use of perspective-taking strategies can be affected by the animacy of the depicted item.