Carnegie Mellon University Annual Progress Report: 2011 Formula Grant

Transcription

1 Carnegie Mellon University Annual Progress Report: 2011 Formula Grant Reporting Period July 1, 2012 June 30, 2013 Formula Grant Overview The Carnegie Mellon University received $943,032 in formula funds for the grant award period January 1, 2012 through December 31, Accomplishments for the reporting period are described below. Research Project 1: Project Title and Purpose Correlated Structure in Motor Cortical Populations Motor control is one of the most important tasks the brain performs, and disorders of motor control affect millions of people. Although a wealth of psychophysical studies have led to good descriptions of the phenomenological processes underlying motor control and adaptation, the neural implementations of these processes are not well understood. One problem is that motor control is inherently a neural population phenomenon: movements are generated by groups of neurons that must work in a coordinated fashion to produce precisely timed muscle activation patterns. Using braincomputer interfaces, we will study how various features of the motor task act to shape the correlation structure of cortical population activity. Anticipated Duration of Project 1/1/ /31/2014 Project Overview Volitional motor control is inherently a neural population phenomenon: to generate movements, neural activity from collections of neurons across multiple brain areas must be coordinated to result in precisely timed muscle activation patterns. This coordination is expressed by statistical dependencies in the tuning of groups of neurons, the so-called signal correlation, which arises from network constraints such as common inputs into groups of neurons. In motor control, these common inputs relate to the cognitive and behavioral factors underlying movement generation. By studying how signal correlations relate to various features of the task, like feedback or redundancy, we can probe how these parameters coordinate population activity in motor cortex and, ultimately, shape motor planning. This project is a coupling of experimental and computational approaches to characterize the flexible correlation structure of motor neuronal populations. We will have monkeys implanted with chronic multi-electrode recording arrays perform a combination of motor tasks including Carnegie Mellon University 2011 Formula Grant Page 1

2 arm reaching and brain-computer interface (BCI) cursor control. Data from these tasks will be used to build a statistical model that fits the correlation structure as a function of both volitionally controllable driving inputs and task-dependent sensory feedback. Ultimately, the lessons we learn from the formulation of these models will improve our understanding of the cognitive and computational principles of motor control, and unite the neural encoding of movement with behavioral theories of motor control. Specific Aim 1: Dissociate volitional from non-volitional dependencies in correlation patterns. Specific Aim 2: Describe how correlation patterns change as a function of task redundancy. Principal Investigator Steven M. Chase, PhD Assistant Professor Carnegie Mellon University 115 Mellon Institute 4400 Fifth Ave Pittsburgh, PA Other Participating Researchers Andrew Schwartz, PhD employed by the University of Pittsburgh Expected Research Outcomes and Benefits We anticipate several potential outcomes and benefits resulting from this study. (1) This research has direct implications for improving neural prosthetic devices, which have the potential to improve the quality of life for a substantial population of patients living with neurological movement disorders. Results from this study will be leveraged to current, ongoing clinical trials. (2) Robotic controllers generally lack the flexibility and robustness exhibited in physiological motor control. By furthering our understanding of the basic mechanisms of motor control, our findings could improve the design and performance of general autonomous control systems across a wide variety of applications. (3) Graduate students supported by this grant will be extensively cross-trained in both computational and experimental approaches to systems neuroscience, placing them at the forefront of a rapidly growing field. (4) Methodologies developed during this study will be directly incorporated into the classes taught by the Principal Investigator. Carnegie Mellon University 2011 Formula Grant Page 2

3 Summary of Research Completed Milestone for reporting period: Complete Aim 1 experiments, start experiments outlined in Aim 2. Progress: 1: Experimental training progress (Specific Aims 1 & 2) Over the past fiscal year, we have continued to build the infrastructure necessary to complete the experimental portions of this work. We have acquired one non-human primate (NHP), and have started training him in the behavioral tasks that he will need to learn to perform before he can be implanted with chronic multi-electrode recording arrays and trained to perform the braincomputer interface tasks that are the centerpiece of this work. Specifically, (1) we have transitioned the subject from ad-lib water to water-reinforcement, in which the subject must interact with the researchers to attain his daily water ration, (2) we have completed chair training, where the subject learns to transition from his home cage to the NHP restraint chair in which he will sit during experiments, and (3) we have begun training on the center-out arm reaching task. This last task involves familiarizing the subject with our virtual reality paradigm, in which the subject sees targets on a 3D stereoscopic monitor and must learn to move his arm (represented by a green cursor in the 3D monitor) to intersect with those targets. Arm position is monitored with an infrared emitting marker attached to the back of the subject s wrist and tracked by a bank of 3 cameras that triangulate its position. Currently, the subject is undergoing marker familiarization training so that he learns not to pick at the marker and change its position on his arm. 2: Off-line analysis of volitional correlation patterns (Specific Aim 1) We have performed some off-line analysis of data previously collected by the Schwartz lab in another context that we are now analyzing in relation to specific aim 1. In that experiment, an NHP performed a 1D brain-computer interface control task, in which the radius of a ring on a computer screen was made proportional to the firing rates of one or more neurons. The goal of the subject was to modulate his firing rates to make the radius of the ring match either a large target ring or a small target ring. We are now investigating the correlation patterns we observe across the range of neurons recorded during this task, some of which were directly in control of the ring, but most of which were not. We are looking to see if the correlation patterns we observe in this task are the same or different as the correlation patterns we observe in more standard center-out reaching tasks. This will inform us of the range of volitionally controllable correlation patterns the subject is capable of displaying to solve the task, and inform us as to whether the standard tasks are sufficient for capturing the range of correlation patterns expressible by the population. Our preliminary analyses indicate that neither the center-out task nor the 1D ring task adequately capture the range of correlation patterns the subject is capable of displaying. To test this, we performed principal components analysis (PCA) on data collected either during the braincontrolled rings task or a natural reaching center-out task, and quantified how particular correlation patterns (from firing rates to one of the targets) was captured by the principal Carnegie Mellon University 2011 Formula Grant Page 3

4 components from each task. Figure 1 shows that when the principal components are computed from data within the same task, a small number of components can explain the majority of the variation in the population (on-diagonal plots). However, when data from the alternate task is used, it does not tend to capture the range of correlation patterns in the low-order principle components. These data indicate that exploring the range of volitional dependencies in neural data will require a wide range of experimental tasks. 3: Off-line analysis of neural adaptation in response to task redundancy (Specific Aim 2) Lastly, we are also performing off-line analysis of data previously collected by the Schwartz lab in relation to specific aim 2. In that experiment, an NHP performed a center-out reaching task under brain-control, where the cursor movement was controlled solely by the modulation of neural activity in primary motor cortex. The subject first performed the center-out reaching task in 3D (that is, he moved the cursor from the center of the workspace to targets distributed on a sphere in 3D); he then used the same group of neurons to perform the task in 2D (where he moved the cursor from the center of the workspace to targets distributed around a circle in 2D). The only difference between the tasks is that while in the 3D the subject could move the cursor anywhere within the 3D virtual space, during the 2D task the z-component of the decoded trajectory was set to zero, so that the cursor was constrained to lie in the xy-plane. Thus, the 2D and 3D tasks were almost exactly the same, with the exception that the 2D task had one less decoding degree-of-freedom than the 3D task, and thus had one extra dimension of redundancy. We have analyzed the neural tuning to target direction for the overlapping set of targets common to the two tasks. Interestingly, we find that the dynamic range of the neurons (that is, the difference between the maximum firing rate to the targets and minimum firing rate to the targets) increases in the 2D task relative to the 3D task. Examples of this phenomenon are given in Fig. 2; data from all cells on one day are given in Fig. 3A. We also see that when we consider the dynamic range of the cells to all targets in the 3D task, not just the ones in the xy-plane, that the dynamic range between the 2D and 3D tasks is roughly equal (Fig. 3B). Our interpretation of this data is that the dynamic range of each neuron appears to be reallocated to be fully spanned by the targets in use in any given task. If this finding holds true for more subjects, it would indicate that neurons dynamically increase their information capacity to encode the current task, and therefore more-redundant tasks would be coded with higher information resolution than lessredundant tasks. Carnegie Mellon University 2011 Formula Grant Page 4

5 Figure 1: We observe a diversity of correlation patterns in different tasks. Each pie chart represents how well a set of 7 principal components (PCs) accounts for the variation observed in neural population activity. Color denotes PC number. Along the diagonal, the task used to compute the PCs is the same as the task used to assess how well those PCs account for the data. For the rings task, 4 PCs account for over 99% of the variance in neural population activity (upper-left), for the center-out task, 4 PCs account for 91% (lower-right). On the off-diagonal, the task used to compute the PCs is different from the task used to assess how well they account for the data. In each case, much of the data is only captured by higher-numbered PCs: 23% of the center-out variance is accounted for by PCs 5-7 of the rings task (upper-right), while 30% of the rings variance is accounted for by PCs 5-7 of the center-out task (lowerleft). Carnegie Mellon University 2011 Formula Grant Page 5

6 Figure 2: Examples of tuning curve rescaling. Above are plotted the planar tuning curves (firing rate as a function of the planar target direction in degrees) for 15 example neurons (out of a total of 29) recorded during two tasks: brain-computer interface reaching to targets during a 2D (red) or 3D (black) center-out task. For the majority of neurons, the average firing rates show an increased dynamic range (difference between maximum and minimum rates) in the 2D task relative to the 3D task. Particularly good examples are indicated with a *. Carnegie Mellon University 2011 Formula Grant Page 6

7 Figure 3: Neural dynamic range increases to encompass full task. A. Each dot represents the dynamic range from one neuron to targets in the xy plane when presented during the 2D task (y-axis) or when presented during the 3D task (x-axis). The dynamic range to the same target set increases when they are the only targets relevant for the task (p=.002). B. 2D dynamic range plotted as a function of the full 3D dynamic range. The results are not statistically different (p=.49). The results indicate that during the 2D task, the entire dynamic range measured during 3D is applied to targets that lie in the plane. Research Project 2: Project Title and Purpose Non-invasive Optical Imaging of Perceptual Learning and Development The reliability and consistency of ordinary sight and hearing makes it natural to presume that perceptual systems are hard-wired and stable. Instead, however, they are highly dynamic and adapt flexibly to allow perceivers to discover regularities in the environment. In fact, over time perceptual expertise develops such that the brain s response to some classes of highly significant stimuli (faces, written words, speech) is markedly distinct. Our ultimate objective is to understand the learning mechanisms that serve the development of perceptual expertise to better understand developmental disorders (autism, dyslexia) and brain injuries that affect perception and to engineer devices to improve perception among those with impairments. Anticipated Duration of Project 1/1/ /31/2014 Project Overview Our ultimate goal is to understand how perceptual systems are shaped by experience to develop perceptual expertise for some classes of stimuli. Humans exhibit such expertise for faces and Carnegie Mellon University 2011 Formula Grant Page 7

8 native-language speech sounds and, later with developing literacy, for printed words. Human perceptual expertise for recognizing and categorizing these stimuli well exceeds the capacity of even the most sophisticated software for face and speech recognition. Understanding the learning mechanisms involved with extracting perceptual regularity from an inherently noisy and variable environment will provide insight about how to improve automatic machine recognition systems. It will also inform how to remediate perceptual problems arising from brain injury and developmental disorders and how to build effective rehabilitation programs for individuals with perceptual impairments. The specific long-term aim is to address the development of perceptual expertise among preschool aged children, a developmental window during which perceptual systems are thought to be highly malleable and during which time developmental disorders that impact perceptual expertise (autism, specific language impairment) tend to be discovered. The present research is unique and innovative because it allows for simultaneous measurement of brain and behavioral responses in young children who are unable to participate in many other forms of neuroimaging research, thus allowing us to probe the development of perceptual expertise. In the present project, we test the hypothesis that adaptation of an optical neuroimaging signal will serve as a more sensitive measure of children s phonetic representations than their behavioral responses and 2) that acoustic context will further shape these representations. At a broad level, achieving our aims will set the stage for Carnegie Mellon University s faculty to focus its research expertise toward innovative new approaches to studying the development of perceptual expertise and its health implications. Principal Investigator Lori L. Holt, PhD Professor Carnegie Mellon University 5000 Forbes Avenue Pittsburgh, PA Other Participating Researchers Kai-Min Chang, PhD; Marlene Berhmann, PhD; David Plaut, PhD; David Rakison, PhD; Erik Thiessen, PhD; Anna Fisher, PhD; Nathan Urban, PhD; Michael Tarr, PhD employed by Carnegie Mellon University Theodore Huppert, PhD employed by the University of Pittsburgh Expected Research Outcomes and Benefits Perceptual learning is a robust phenomenon, measurable throughout the lifespan in humans and other species that is thought to support a variety of basic cognitive, perceptual and language functions. It is important because changes in the way that perceptual input is processed and represented impact all subsequent processing at higher levels. Understanding brain function as it relates to developing perceptual expertise for particular classes of stimuli will allow us to generate models of how perceptual learning can be harnessed to improve perceptual processing Carnegie Mellon University 2011 Formula Grant Page 8

9 (such as training physicians to better detect tumors in visually-noisy scans) and remediate everyday perception when perceptual systems go awry. Deficits in perceptual learning are observed across a variety of brain disorders including schizophrenia for auditory processing, autism for faces and speech, specific language impairment and dyslexia for spoken language and words, and in response to traumatic brain injury, stroke, and seizure. Determining how to engineer systems to remediate brain disorders affecting perceptual processing requires an understanding of the relationship of brain function to perceptual learning and how perceptual expertise for specific classes of stimuli develops. Neuroimaging tools that can be used effectively to study the developing brain are necessary in this endeavor. The present project exploits functional near-infrared spectroscopy (fnirs) as a neuroimaging technique suitable for use with children and therefore significant for measuring brain development. By understanding how the brain changes with development of perceptual expertise, we believe we can better understand the causes of perceptual symptoms for brain disorders. Summary of Research Completed Milestone for reporting period: Continue work toward repurposing research space; Hire technician; Accept shipment of NIRS system from TechEn; Install NIRS system; Begin to test device with assistance of TechEn engineers under the supervision of Dr. Theodore Huppert, Co- PI, who has extensive experience with the system; Continue working on suitability of research space for ultimate developmental testing; Begin to train investigators in fnirs; Pilot testing of materials with adult listeners. Progress: We have achieved the following aims: 1) In the last two project years we have paved the way for near-infrared spectroscopy (NIRS) research with adults and children at Carnegie Mellon University. Our ultimate goal authorize NIRS as minimal risk, as it is on other campuses internationally (and as are other neuroimaging techniques like functional magnetic resonance imaging, fmri), so that we may involve children at the Carnegie Mellon Children s School in our experiments. This process has involved meetings with the directors of the IRB and the university compliance office. To support our efforts, we invited a prominent NIRs researcher, Dr. Janet Werker of University of British Columbia, to meet with our IRB representatives and answer their questions. This Spring, a proposal was submitted for consideration to the Carnegie Mellon University Institutional Review Board. 2) Over the course of the last year we have interacted with TechEn, the manufacturer of the NIRs equipment, to develop a price quote and initiate construction of the custom-made NIRs system. The custom-built system shipped to Carnegie Mellon in late Spring ) We negotiated with administrative bodies on campus to secure space for the new equipment. In this project year we secured the space and adjusted it to suit the needs of the NIRS system. 4) On May 13, 2013 representatives of TechEn successfully installed and tested the equipment. 5) In tandem with these necessary administrative goals, we have made scientific progress. In this project year we undertook a collaborative project with Carnegie Mellon University s Carnegie Mellon University 2011 Formula Grant Page 9

10 Entertainment Technology Center (ETC). The ETC team worked with us to develop phonetic learning software embedded in a videogame. This will provide an experimental paradigm suitable for testing our hypothesis that adaptation of an optical neuroimaging signal will serve as a more sensitive measure of children s phonetic representations than their behavioral responses. Carnegie Mellon University 2011 Formula Grant Page 10