Kenneth Kreutz-Delgado - US grants

This Presidential Young Investigator Award will support Dr. Kreutz-Delgado's development of a real-time testbed for sensor- based control of a dual-arm robotic manipulation system. His goal is exploration of robotic assembly and servicing principles for flexible manufacturing and for autonomous robotic systems. His current research, begun at the Jet Propulsion Laboratory, involves application of a spatial operator algebra to real-time task-level planning and use of neural-network approaches for manipulator and camera coordination.

This is the first year of a three-year continuing award. This research investigates task-level planning for autonomous agents, such as mobile robots, that function in an uncertain environment. These robots typically have very approximate, inaccurate, or minimal models of the environment. For example, although the geometry of its environment is crucial to determining its performance, a mobile robot might only have a partial, or local "map" of the world. The research will investigate an approach whereby the robot attempts to acquire the necessary information about the world by planning a series of experiments using the robot's sensors and actuators, and building data structures based on the robot's observations of these experiments. A key feature of this approach is that the experiments the robot performs should be driven by the information demands of the task. This research develops (1) a theory of sensor interpretation and task-directed planning using perceptual equivalence classes, intended to be applicable in highly uncertain or unmodelled environments, such as for a mobile robot; (2) algorithmic techniques for modeling geometric constraints on recognizability, and the building of internal representations (such as maps) using these constraints; (3) explicit encoding of the information requirements of a task using a lattice (information hierarchy) of recognizable sets, which allows the robot to perform experiments to recognize a situation or a landmark; and (4) synthesis of robust mobot programs using the geometric constraints, constructive recognizability experiments, and uncertainty models imposed by the task. This research will also implement the theory and test it on mobile robots in the laboratory.

9415483 Skalak This award is to conduct the annual NSF Design and Manufacturing Grantees Conference. The conference will involve researchers from the Division of Design, Manufacture, and Industrial Innovation and the Division of Engineering Education and Centers, both located within the Directorate for Engineering, and the Division of Information, Robots and Intelligent Systems, located within the Directorate for Computer and Information Science and Engineering. The conference ensures that the individual researchers are informed about the ongoing activities of their colleagues. An elimination of duplication of their efforts may be achieved and a degree of cooperation may result from this activity. An overall improvement of efficiency of the research activity could be expected. In addition, the conference program organization allows for ample time to discuss manufacturing research in detail with the collective research community at the meeting, with feedback to and input from the National Science Foundation. Finally personal contacts between the grantees and program directors in the Divisions should contribute to clarifying many current issues in their work. The aims of the conference are to enhance communication between researchers, encourage joint research, promote synergism, and develop networks for technical interchange. Attendees to the conference gain an early access to the information disseminated.

9422069 Jain For the past decade, the importance of visual computing has increased exponentially. Visual computing, which embraces processing, interpreting, modeling, assimilating, storing and synthesizing visual information, now plays a pivotal role in many fields. These include such subjects as: virtual reality, multimedia, robotics, computer-human interaction, scientific visualization and communication. This award is to purchase equipment for supporting a number of ongoing research projects that are committed to this important field. The goal of these projects is to improve visual information computing through innovative experimental research. The equipment which include two visualization systems, one real-time image processing subsystem, and a number of visual sensors will be dedicated to support four individual research projects each addressing a different aspect of visual computing, namely Visual Information Assimilation, Visual Interaction through Gesture Recognition, Modeling and Design of Optoelectronic Visual Information Processors, and Physics-based Visualization of Multipedal Walking Systems. ***

he central active challenge we are constantly addressing in daily life is to correctly assess the intent of others ('What is she trying to do? ...') and the import of sensory events ('What - good or bad - may happen now? ...') based on active perception ('It looks to me like she is trying to ...') and retrieved associations (''And she was the one who ...'). The corresponding problem for cognitive neuroscience is to identify, ideally from non-invasive brain activity recordings, those patterns of distributed brain activity that accompany and support active human cognition and behavior. This problem has two parts: First, -What patterns of distributed brain dynamics follow from, accompany, and predict specific world events and subject behavior? -To fully understand the experience and behavior of subjects in performing a given task, we must take into account both the import of each task event to the subject and the intent of each of behavioral event. These factors cannot be known directly, but they may be accurately guessed or inferred, in many cases, from detailed recordings of subject behavior and from the specific historical context in which each recorded environmental or behavioral event occurs. In the case of electroencephalographic (EEG) and/or magnetoencephalographic (MEG) signals recorded non-invasively from the human scalp, a second part of the problem remains -Which brain areas generate the identified signal patterns?'

The usual approach to analyzing electromagnetic scalp data has been to separate recorded events and behavior into a few simple categories, to average the recorded brain dynamics time locked to each event category, and then to apply physical inverse source estimation methods to scalp maps of peaks in the resulting averages. This project will explore using new machine learning methods, including advanced independent component analysis (ICA) and sparse Bayesian learning (SBL) methods, to jointly model the recorded task event, subject behavior, and brain dynamic data recorded in a complex learning task. The project has two goals: First, to identify patterns in unaveraged EEG and/or MEG data that reliably accompany subject behavior in specific contexts, and second to determine the exact areas of the subject's cortical mantle that locally synchonize their electromagnetic activities to produce the identified scalp patterns. If successful, the project will enhance the value of noninvasive electromagnetic brain imaging for identifying and measuring, with high temporal and spatial resolution, complex, distributed patterns of locally synchronous cortical activity that support active human cognition.

Abstract
This research examines theoretical, algorithmic, and computational issues that arise in signal processing problems where there is a need to compute sparse solutions. There are numerous signal-processing applications where sparsity constraint on the solution vector naturally arises. Brain imaging techniques such as MEG and EEG, sparse communication channels with large delay spread, high-resolution spectral analysis, direction of arrival estimation and compressed sensing are a few examples. The generalization and extension of the sparse Bayesian learning (SBL) techniques considered in this research will broaden the application domain and provide a very powerful complement to the existing maximum a posteriori (MAP) methods commonly used and in some cases even surpass them.
The investigators study extensions and generalizations of the sparse source recovery problem to greatly broaden the application domain. A key consideration in the work is developing a rigorous framework to deal with dependency in the sparsity framework. Motivated by applications with sparse but local structure, the research considers intra-vector dependency in the single measurement case, as well as intra-vector dependency as required in the multiple measurement contexts, among others. The research also includes the development of connections between multi-user communication theory and the sparse signal recovery problem to shed light on the stability with which sparse signal recovery is possible and to develop an understanding of the limits of suboptimal source recovery methods. To deal with non-stationary environments, the research develops on-line adaptive algorithms that exploit the inherent sparse structure of the application. The research also includes evaluation of the resulting algorithms in several important application domains.

Level of Effort Statement
At the recommended level of support, the PI and co-PI will make every attempt to meet the original scope and level of effort of the project.

Intellectual Merit: This project aims at combining cognitive and computational neuroscience, neuroengineering and system identification towards a transformative understanding of the way distributed brain dynamics interact with motor activity in humans. 3-D body and limbs movement kinematics, eye movements and electroencephalographic (EEG) spatiotemporal brain data will be recorded simultaneously during motor control and adaptation in healthy and Parkinson?s disease patients. In particular, altered and real world motor tasks will be simulated in 3-D immersive virtual reality technology with force feedback robots providing proprioceptive interaction and feedback. Cognitive, behavioral and kinematics data will constrain the design of large-scale computational models of motor control and adaptation based on known anatomy and physiology of the basal ganglia. Neuromorphic engineering will guide the design of mobile embedded computational systems for real-time emulation of the brain-body models and closed-loop sensory-motor control for Parkinson?s patients. We expect that the development of new machines for neuro-rehabilitation will result in a threefold synergetic interaction between engineering and neuroscience: human-machine interactions will transform the notion of movement control and provide new contexts to study embodied cognition that will benefit neuroscience; in turn, new knowledge in neuroscience and motor control will accelerate the development of adaptive machines for rehabilitation and/or enhancement. Finally, comprehensive and predictive mathematical models of motor control implemented in neuromorphic hardware are expected to lead to new intelligent neuroprosthetic tools.

Broader Impact: Outcomes of this research will contribute to the system-level understanding of humanmachine interactions and motor learning and control in real world environments for humans, and will lead to the development of a new generation of wireless brain and body activity sensors and adaptive prosthetics devices. This will advance our knowledge of human-machine interactions, stimulate the engineering of new brain/body sensors and actuators, and have a direct influence in diverse areas where humans are coupled with machines, such as brain-machine interfaces, prosthetics and telemanipulation. We anticipate that the confluence of cognitive and computational neuroscience, control theory and wearable, unobtrusive bioinstrumentation will provide novel non-invasive approaches or the treatment and neuro-rehabilitation of Parkinson?s disease and will potentially transform our understanding of brain/body interactions. The project draws graduate and undergraduate students across divisions and in the NSF Temporal Dynamics of Learning Center (TDLC) and Institute of Neural Computation (INC) at UCSD participating in interdisciplinary engineering and neuroscience aspects of the design, optimization, and training of largescale neuromorphic systems and their human interfaces. Through outreach channels on campus supported by the TDLC and the NSF Research Experience for Undergraduates (REU), the program will actively pursue increased participation in research and education of the next generation of scientists and engineers.

This project, called the Cognitive Hardware And Software Ecosystem Community Infrastructure (CHASE-CI), will build a cloud of hundreds of affordable Graphics Processing Units (GPUs), networked together with a variety of neural network machines to facilitate development of next generation cognitive computing. This cloud will be accessible by 30 researchers assembled from 10 universities via the NSF-funded Pacific Research Platform. These researchers will investigate a range of problems from image and video recognition, computer vision, contextual robotics to cognitive neurosciences using the cloud to be purpose-built in this project.

Training of neural network with large data-sets is best performed on GPUs. Lack of availability of affordable GPUs and lack of easy access to the new generation of Non-von Neumann (NvN) machines with embedded neural networks impede research in cognitive computing. The purpose-built cloud will be available over the network to address this bottleneck. PIs will study various Deep Neural Network, Recurrent Neural Network, and Reinforcement Learning Algorithms on this platform.