Byron M. Yu, Ph.D. - US grants

DESCRIPTION (provided by applicant): Brain-computer interfaces (BCI) can assist paralyzed individuals and amputees by translating their neural activity into movements of a BCI plant, such as a computer cursor or prosthetic limb. For many years, the field sought offline decoders that could best map neural activity to arm movements. It has become increasingly recognized that designing an effective online, closed-loop decoder is quite a different challenge. A key difference is that, in a closed-loop setting, the subject receives sensory feedback about the state of the BCI plant and can compensate for errors by generating new neural activity patterns. To engineer clinically-viable, closed-loop BCI systems, many fundamental questions about the neural underpinnings of their performance must be answered. Can subjects generate arbitrary neural activity patterns to compensate for errors? Do subjects form an internal model of the BCI plant to achieve proficient control in the presence of noisy, delayed feedback? Do subjects exploit the redundancy inherent in the mapping from neural activity to BCI plant kinematics to maximize control accuracy? A critical roadblock for answering these questions is the lack of an appropriate statistical framework to rigorously analyze closed-loop BCI data on a timestep-by-timestep basis. We propose to develop such a framework inspired by control theory, in close conjunction with novel closed-loop BCI experiments. We will train non-human primates to perform dextrous control of a BCI cursor using neural activity recorded in primary motor cortex with chronic, multi-electrode arrays. We will test the hypothesis that BCI learning depends on constraints imposed by the underlying neural circuitry. In parallel, we will develop and validate algorithms to explain the observed, high-dimensional neural activity at each timestep by accounting for the sensory feedback, subject's internal model of the BCI cursor, and behavioral task goals. We will then leverage the developed algorithms to investigate whether subjects can exploit neural redundancy during BCI control. Broader Impact: We envision five areas of broader impact. First, BCI systems promise to dramatically improve the quality of life for disabled patients. Clinical trials are ongoing, so opportunities exist to translate our research directly and in the near term into clinical practice. Second, our understanding of the neural basis for arm movement control is still incomplete, in large part because the system is so complex. BCIs provide a simplified motor control system, where a well-defined relationship exists between neural activity and movement. As such, BCIs provide a novel experimental testbed to investigate the neural mechanisms of motor control and learning. Third, the statistical framework we develop may be applicable to the study of feedback control systems in other domains. Fourth, with the advent of large-scale neural recordings, systems neuroscience is becoming a far more quantitative field. The next generation of researchers must be well-versed in computational and biological principles. We believe that our collaboration provides an excellent dual-training environment for our students and postdocs. Fifth, our research discoveries can directly feed into our classroom teaching. Yu teaches Neural Signal Processing at CMU and Batista teaches Control Theory in Neuroscience at Pitt;both are annual, graduate-level courses. Intellectual Merit: In the last decade, several groups have demonstrated compelling proof-of- concept laboratory demonstrations of closed-loop BCI control. For clinical translation, one of the major challenges is to improve the performance and robustness of BCI systems. To make this leap, we believe that it is critical to rigorously study existing systems to understand i) why some BCI decoders work better than others, ii) to what extent we can depend on the subjects'ability to learn, and iii) the neural strategies adopted by the subjects for proficient control. There is a long-overdue need for a general statistical framework for dissecting closed-loop BCI data, which we propose to develop. Discoveries enabled by the developed methods will help us and others in the field to design high-performance, clinically-viable BCI systems that allow the subject to quickly reach and maintain a high level of proficiency.

DESCRIPTION (provided by applicant): Brain-computer interfaces (BCI) can assist paralyzed individuals and amputees by translating their neural activity into movements of a BCI plant, such as a computer cursor or prosthetic limb. For many years, the field sought offline decoders that could best map neural activity to arm movements. It has become increasingly recognized that designing an effective online, closed-loop decoder is quite a different challenge. A key difference is that, in a closed-loop setting, the subject receives sensory feedback about the state of the BCI plant and can compensate for errors by generating new neural activity patterns. To engineer clinically-viable, closed-loop BCI systems, many fundamental questions about the neural underpinnings of their performance must be answered. Can subjects generate arbitrary neural activity patterns to compensate for errors? Do subjects form an internal model of the BCI plant to achieve proficient control in the presence of noisy, delayed feedback? Do subjects exploit the redundancy inherent in the mapping from neural activity to BCI plant kinematics to maximize control accuracy? A critical roadblock for answering these questions is the lack of an appropriate statistical framework to rigorously analyze closed-loop BCI data on a timestep-by-timestep basis. We propose to develop such a framework inspired by control theory, in close conjunction with novel closed-loop BCI experiments. We will train non-human primates to perform dextrous control of a BCI cursor using neural activity recorded in primary motor cortex with chronic, multi-electrode arrays. We will test the hypothesis that BCI learning depends on constraints imposed by the underlying neural circuitry. In parallel, we will develop and validate algorithms to explain the observed, high-dimensional neural activity at each timestep by accounting for the sensory feedback, subject's internal model of the BCI cursor, and behavioral task goals. We will then leverage the developed algorithms to investigate whether subjects can exploit neural redundancy during BCI control. Broader Impact: We envision five areas of broader impact. First, BCI systems promise to dramatically improve the quality of life for disabled patients. Clinical trials are ongoing, so opportunities exist to translate our research directly and in the near term into clinical practice. Second, our understanding of the neural basis for arm movement control is still incomplete, in large part because the system is so complex. BCIs provide a simplified motor control system, where a well-defined relationship exists between neural activity and movement. As such, BCIs provide a novel experimental testbed to investigate the neural mechanisms of motor control and learning. Third, the statistical framework we develop may be applicable to the study of feedback control systems in other domains. Fourth, with the advent of large-scale neural recordings, systems neuroscience is becoming a far more quantitative field. The next generation of researchers must be well-versed in computational and biological principles. We believe that our collaboration provides an excellent dual-training environment for our students and postdocs. Fifth, our research discoveries can directly feed into our classroom teaching. Yu teaches Neural Signal Processing at CMU and Batista teaches Control Theory in Neuroscience at Pitt; both are annual, graduate-level courses. Intellectual Merit: In the last decade, several groups have demonstrated compelling proof-of- concept laboratory demonstrations of closed-loop BCI control. For clinical translation, one of the major challenges is to improve the performance and robustness of BCI systems. To make this leap, we believe that it is critical to rigorously study existing systems to understand i) why some BCI decoders work better than others, ii) to what extent we can depend on the subjects' ability to learn, and iii) the neural strategies adopted by the subjects for proficient control. There is a long-overdue need for a general statistical framework for dissecting closed-loop BCI data, which we propose to develop. Discoveries enabled by the developed methods will help us and others in the field to design high-performance, clinically-viable BCI systems that allow the subject to quickly reach and maintain a high level of proficiency.

Movements are inherently variable: one never throws a dart or a basketball in exactly the same way twice. On the face of it, this variability in behavior is detrimental to performance, preventing one from consistently hitting the bull's-eye or making the basket. However, computational theories posit that motor variability may also serve a functional role, enabling exploration and learning of more efficient movements. This creates an intriguing duality: while variability should be minimized for short-term motor performance (to act reliably), it should be maximized for long-term performance (to promote learning). During practice, variability might be useful for developing motor skill. When it's game time, however, variability should be suppressed to the greatest extent possible. Might the central nervous system set the amount of variability in a context-appropriate fashion? This study will investigate the neural correlates of motor variability and establish the connections between neural variability, behavioral performance, and learning.

Neural variability lies at the heart of several theoretical computational models, from implementations of probabilistic computation to Hebbian learning rules. Although the importance of variability has been well recognized, the structure and regulation of neural variability within the central nervous system is not well understood. This project coordinates a program of experiments and new analytical techniques to examine the structure of neural variability in the motor system. It seeks to establish, first, how variability depends on behavioral demands, and second, how variability impacts learning. To achieve this, many neurons of the motor and premotor cortices will be studied simultaneously during performance of demanding behaviors. By studying two distinct areas in the motor pathway, the impacts of noise on motor planning and execution can be examined separately. Furthermore, population recordings can be leveraged to decompose variability into three conceptually distinct components: (1) variability that is related to the task (signal variability), (2) trial-to-trial variability shared among neurons, and (3) private variability within each neuron. The investigators will explore how variability of each type is modulated by task context and learning. These decompositions will yield insight into the mechanisms of variability generation during performance.

This project is funded by Integrative Strategies for Understanding Neural and Cognitive Systems (NSF-NCS), a multidisciplinary program jointly supported by the Directorates for Computer and Information Science and Engineering (CISE), Education and Human Resources (EHR), Engineering (ENG), and Social, Behavioral, and Economic Sciences (SBE). Even basic perception of the world is not as simple as light coming into the eyes or sound coming into the ears. Rather, perception involves combining the incoming sensory information with cognitive processes such as past experiences, knowledge about the world, and personal tendencies. In other words, two people observing the same events (i.e., receiving the same sensory information) can arrive at different interpretations of what is happening in the environment. How the brain combines sensory information with these cognitive processes, and where this occurs in the brain, is incompletely understood. The key innovation of this project is to use a brain-computer interface (BCI) to tease apart which aspects of the brain's activity are sensory versus cognitive and how the two are combined in the brain to produce perception of the world. BCIs are widely-known for their ability to help paralyzed patients and amputees by allowing them to move a computer cursor or robotic arm simply by thinking about moving. Few studies have used BCIs as an experimental tool to understand sensory areas of the brain, as this project seeks to do. This work is likely to lead to a deeper understanding of how we perceive the world, as well as insights into how BCI can be used to help treat psychiatric disorders and recover function after injury. Furthermore, the investigators are developing BCI-based lab exercises for undergraduate courses, training researchers to become well-versed in experimental and computational neuroscience, and involving undergraduates, including women and underrepresented minorities, in the research.

This project focuses on visual area V4, which is known to be a crossroads for sensory and cognitive processes during visual perception. To dissect what aspects of neural activity are sensory versus cognitive, the investigators train animal subjects to volitionally modulate their V4 activity. The BCI provides subjects with moment-by-moment auditory feedback of their V4 activity. This project assesses what aspects of V4 activity can be volitionally (i.e., cognitively) modulated, how volitional modulation of V4 activity affects visual perception, and how malleable is the interaction between V4 and another brain area (prefrontal cortex) during visual perception. The key advantage of using BCI for this study is that it allows the investigators to challenge the subjects to produce particular patterns of neural activity. The investigators can specify in the BCI which patterns of activity yield a reward. This technique allows them to assess what aspects of the neural activity can be volitionally controlled by the animal (i.e., cognitive), and what aspects are hard-wired to the outside world (i.e., sensory). The applications of this BCI paradigm are extremely broad, and can be used to study other sensory, cognitive, and motor systems.

Cognition and behavior unfold over time. The temporal aspects of thought and action reflect, at least in part, the temporal evolution of the activity of the populations of neurons that control them. It has long been hypothesized that the time course of neural activity arises from dynamical constraints imposed by the underlying neural circuitry. However, this has been difficult to show experimentally because it requires the ability to finely perturb neural activity in varied ways. Here, we propose to employ a brain-computer interface (BCI) paradigm to study neural dynamics. A BCI enables us to perturb neural activity by harnessing the animal's volitional control to drive the activity of a population of neurons into configurations that we specify. In this way, we can perform causal tests of dynamical constraints and their relation to behavior. We will study dynamical constraints imposed by motor preparation using multi-neuronal activity recorded in the motor cortex of macaque monkeys. We hypothesize that dynamical constraints exist in the motor cortex, and that these constraints are shaped during motor preparation to drive arm movements. To test these hypotheses, we will first challenge the animals to violate the putative dynamical constraints. Then, we will test the hypothesis that motor preparation sets up dynamical constraints appropriate for the upcoming arm movement. Finally, we will use the BCI paradigm to perturb neural activity during movement preparation to alter and evoke arm movements. Taken together, our proposed work will likely lead to a richer understanding of how networks of neurons give rise to population dynamics, and how those dynamics relate to neural computation and behavior.

Project Summary Behavior and cognition emerge from the coordinated activity of populations of interacting neurons. To change behavior we must change the architecture of the network that drives behavior. What are the rules whereby the activity of populations of neurons can change during learning? If we could observe the neural population dynamics that underlie skill learning, we might be able to facilitate learning. This ability may eventually lead to improvements to neurally- based rehabilitation strategies that can facilitate the recovery from stroke. We will use a neurofeedback paradigm to shape the neural population dynamics that underlie skill learning. To do this, we will record the activity of dozens of neurons in the motor cortex of Rhesus monkey subjects. Animals will control a cursor on a computer screen by generating neural command signals. This is neurofeedback because the animal directly observes a projection of his neural activity, in the form of the movement of the onscreen cursor. This paradigm allows us to study learning simply by perturbing the mapping from neural activity to cursor movement on the screen. Following such a perturbation the animal must discover how to generate new patterns of neural activity that are now appropriate to restore good control of the onscreen cursor. We will ask two linked questions. First, how does the animal learn to generate new patterns of neural activity? And second, can we facilitate that process? Neurofeedback-based learning offers complementary advantages to arm movements for studying the neural population dynamics that accompany learning. Chie?y, only in a neurofeedback paradigm can we be certain that the learning-induced changes in neural activity we observe matter directly for behavior, because only the neurons we record directly impact behavior in this paradigm. We will test classic theories of skill learning, converted into speci?c hypotheses about how neural activity patterns will change throughout the multi-day course of skill learning. If our neurofeedback-based incremental training schemes do facilitate learning, then this research will inform the work of rehabilitation specialists who work with stroke patients. We (and others) believe that if neurofeedback-based therapies are coupled with standard behavioral therapies for stroke, rehabilitation outcomes will improve.

Understanding how different parts of the brain communicate is perhaps the most fundamental question of neuroscience because it is at the heart of understanding all brain functions and disorders. It is of clinical importance because numerous brain diseases - autism, schizophrenia, attention deficit disorder, and many others - are thought to be due to impaired communication among regions of the brain, and attention in particular is impaired in every major neurological disorder. Even though numerous studies have led to understanding of how single neurons respond to flashes of light or simplified visual objects like lines, relatively little work has been directed toward explicitly learning how groups of neurons communicate with each other, and how that communication enables attending to important information and filtering out distractions. The research described in this proposal seeks to reveal how different parts of the brain communicate to support visual perception. The central question addressed is how different parts of the brain communicate to help select the parts of the visual world that warrant focus, and ignore the parts of the world that are distracting. The specific research aims are designed to (1) reveal how communication between visual and prefrontal cortex modulates over time and how those interactions impact behavior (2) develop statistical approaches to optimize the ability to use stimulation to intervene between these brain regions, and (3) apply these methods with microstimulation in prefrontal cortex to modulate visual responses and, in turn, attentional mechanisms. Together, these aims will have important consequences for the understanding of attention, neuronal communication, and interventional approaches to manipulate the nervous system.

Humans are not, by nature, logical creatures. It takes focus to maintain our composure and not let emotions color our judgement. When we can control our emotional state, we can get “in the zone” and perform well. Failing to do so, we’ll “have a bad day”, or just be “off”. Why does this happen? And, in terms of neurobiological mechanisms, how does this happen? Our emotions are regulated by internal states, such as arousal, attention, and motivation, brain-wide modulatory processes that impact neural function related to perception, decision making, and action. What are the neural mechanisms of those interactions? This project will explore the interactions between internal states and cognitive processing in the cerebral cortex. The investigators will leverage their expertise in “brain training” by giving subjects visual feedback about their neural activity so that they are directly aware of their internal states. In this way, they will study whether subjects are able to better regulate their internal states so that they are able to make perceptual judgments and perform motor skills more consistently at a high level of performance. The investigators will also organize workshops to bring together experts in areas related to this project, train researchers to become well-versed in experimental and computational neuroscience, and enhance the participation of undergraduates, women, and underrepresented minorities in the research.

This project involves three integrated research threads. First, the investigators will use multi-electrode recordings in several regions of the cerebral cortex simultaneously to identify brain-wide signatures of internal states and their effect on the communication between cortical areas. Second, they will train subjects to volitionally control their internal states using neurofeedback. Third, they will examine whether subjects can harness their internal states to accelerate learning and improve performance on challenging perceptual and motor tasks. In these studies, they will focus on three types of internal states -- one that guides us in the spatial world around us (spatial attention), one that manages our alertness throughout the day (arousal), and one that aids our effort in focusing on what lies ahead (motivation). They will study how these internal states interact and to what extent they can be volitionally controlled in three areas across the brain: visual area V4, prefrontal cortex, and motor cortex. Together, their work will provide i) a unified account of the impact of multiple internal states on brain-wide neural computations spanning perception and action, and ii) neurofeedback paradigms to enable subjects to harness their internal states for improved performance.

This project is funded by Integrative Strategies for Understanding Neural and Cognitive Systems (NCS), a multidisciplinary program jointly supported by the Directorates for Biology (BIO), Computer and Information Science and Engineering (CISE), Education and Human Resources (EHR), Engineering (ENG), and Social, Behavioral, and Economic Sciences (SBE).

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.