Rajesh P. Rao - US grants

EIA-0130705 -University of Washington-Guang R. Gao-Adaptive Neurally-Inspired Computing: Models, Algorithms, and Silicon-Based Architectures

Rapid advances in silicon technology over the past few decades have allowed digital
devices to achieve ultra-high speeds in numerical computation. However, a majority of
these devices are prone to catastrophic failure when confronted with circumstances
unforeseen at programming time. Endowing such devices with the ability to adapt and
learn from experience is rapidly becoming a problem of fundamental importance in
information technology. We propose a new approach to solving this problem: building
information technology systems based on neurobiological computation and learning.

We intend to achieve this goal by developing computational models of plasticity and information processing in neurons and networks of neurons in selected sensory and motor areas of the brain; testing these models and their corresponding algorithms in software-based simulations, and designing real-time implementations of these algorithms in silicon using synapse transistors and field-programmable learning arrays (FPLAs).

We expect our research to provide a better understanding of computation within neuronal
networks, and to lead to a new generation of adaptive neuromorphic devices that could be
used for a variety of information technology applications, ranging from signal processing
and pattern recognition to ubiquitous computing and robotic control.

The overall goal of this project is to study learning through imitation and endow robots with the ability to learn in this manner. This cross-disciplinary project has two major goals. The first is to create models for imitation learning in robots that combines techniques from Artificial Intelligence and Bayesian machine learning with insights from cognitive and psychological studies of imitation. The project will use these models to develop a humanoid robot that can learn by watching humans perform specific tasks. The second is to determine what characteristics of a humanoid robot can cause human infants and toddlers to imitate it. This will help shed light on the question of whether infants ascribe intentions to robots. The PI will collaborate with cognitive psychologist, Dr. Andrew Meltzoff also from the University of Washington. The project will foster collaboration between students from both of their labs and provide them with training in carrying out interdisciplinary research.

One of the most outstanding problems in science today is how the activities of the ten billion or so neurons in the human brain allow a person to perceive, think, and act in an intelligent and adaptive manner. Knowing the answer to this question would allow the design of radically new technologies with adaptive capabilities that would far outstrip the capabilities of technologies existing today. Recent behavioral and neurobiological experiments have suggested that the brain may rely on probabilistic principles for perception, action, and learning.
The goal of the proposed research project is to develop a rigorous probabilistic framework for neural computation and to test the resulting models in two ways: (1) in collaborative biological experiments, and (2) in applications involving robotics and brain-machine interfaces. Our specific research goals include:
1. Probabilistic Models of Neural Computation: We will develop new models of neural computation based on treating the problems of sensory information processing and action selection as probabilistic inference problems. We will investigate how biological models such as networks of integrate-and-fire neurons can represent probability distributions and how the propagation of neural activities in such networks can implement algorithms for probabilistic (Bayesian) inference of unknown quantities. We will also explore the connections between well-known neurobiological rules governing synaptic plasticity and statistically-derived learning rules.
2. Experimental Validation using Electrocorticographic Studies: Our models of Bayesian inference will be tested by co-PI Ojemann's group in experiments involving electrocorticographic (ECoG) signals recorded from the human brain in consenting patients being monitored in the days prior to brain surgery. Experiments will focus on testing the predictions of our models in tasks involving visual discrimination, recognition, and sensorimotor integration. Results from the experiments will be used to refine existing models and develop new probabilistic models inspired by neurobiological data.
3. Applications in Probabilistic Robotics and Brain-Machine Interfaces: We will test the robustness of our probabilistic models by implementing the corresponding algorithms on an existing humanoid robot in PI Rao's laboratory. We will be focusing primarily on sensorimotor integration and inference of actions for stable control of movements. Simultaneously, we will explore the applicability of our probabilistic models to brain-machine interfaces. The specific goals are to control a cursor on a computer screen and control a 4-degrees-of-freedom robotic arm by probabilistically inferring real and imagined movements from ECoG signals in real time.
The educational component of the project involves interdisciplinary training for one graduate student, research experiences for undergraduates, and curriculum development in the form of a new graduate level course on brain-machine interfaces.

Intellectual Merit: The proposed research represents one of the first interdisciplinary efforts to develop and test a rigorous probabilistic framework for understanding neuronal computation in the brain. Also novel is the application of neurally-inspired probabilistic models to robotics and brain-machine interfaces, two areas that could benefit tremendously from the robustness and adaptability afforded by such models.
Broader Impact: If successful, this research will lead to a new understanding of computation in the brain, offering unique insights into the mechanisms underlying human behavior and cognition. The application to brain-machine interfaces could dramatically improve the quality of life of paralyzed and disabled patients. The grant will enable the training of a graduate student in a multidisciplinary environment. Promising undergraduates, including students from underrepresented groups, will be paired with graduate students, providing valuable research experience for the undergraduates and mentoring experience for graduate students preparing for industrial and academic careers.

0930908
Rao


Brain-machine interfaces (BMIs) are devices that allow a subject to control objects directly using brain signals. Such devices offer the potential to significantly improve the quality of life of locked-in, paralyzed, or disabled individuals by allowing them to communicate via virtual keyboards and control prosthetic robotic devices. The two dominant paradigms for brain-machine interfacing today rely on non-invasive recording from the scalp (EEG) and invasive techniques based on intracortical implants. EEG signals are extremely noisy, thereby limiting the bandwidth of control signals that can be reliably extracted. Intracortical implants on the other hand yield stronger signals but pose serious health risks.

In this proposal, the PI describes a research program for investigating BMIs based on electrocorticography (ECoG), a relatively new technique that involves recording signals subdurally from the brain surface. These signals have much higher signal-to-noise ratio than EEG signal while at the same time, pose lesser risks than techniques that penetrate the brain surface. The proposed research will address the following key issues:

(1) Exploiting high frequency ECoG signals for BMI: Recent work has shown the existence of broad-spectral ECoG changes at high frequencies during movement and imagery. The PI and his team will explore the application of such ECoG modulation for multi-dimensional control in BMIs.
(2) Neural plasticity of local cortical circuits during BMI: The PI's team will investigate the dynamic range of the spectral changes in ECoG and analyze the adaptations that occur due to brain plasticity during BMI control. This will help pave the way for controlling 3 or more degrees of freedom in a BMI from a single control electrode.

(3) Abstraction of control signals: After extended periods of BMI use, many patients report no longer imagining moving a control limb but rather concentrating on the desired result of the BMI task itself. The PI and his team will explore the creation of new cortical communication pathways underlying such abstraction and leverage these new control signals in expanding the bandwidth of the BMI.
(4) Applications of new control signals to novel BMI paradigms: The BMI techniques will be tested using virtual devices such as cursor-driven menu systems for communication as well as more complex robotic systems such as a prosthetic robotic hand and a humanoid robot.
The educational component of the project involves curriculum development, interdisciplinary training for graduate and undergraduate students, and outreach to K-12 students.

Intellectual Merit: The proposed research represents one of the first efforts to exploit ECoG and the brain's plasticity to build BMIs that can control devices with large degrees of freedom. The study of abstraction of control signals and its application to robotic BMIs is also novel.
Broader Impact: If successful, this research will lead to new ECoG-based BMI systems that will surpass the abilities of current BMIs by relying on the brain's ability to adapt to novel control scenarios and leveraging the large-scale population-level electrical activity measured by ECoG. The project will enable the training of graduate students in a multidisciplinary environment. Promising undergraduates, including students from underrepresented groups, will gain valuable research experience in preparation for industrial and academic careers. A K-12 outreach effort will enable students from local area schools to visit the laboratories of the PIs and gain hands-on experience in the emerging field of brain-machine interfaces.

Humans are extremely adept at learning new skills by watching and imitating others. Attempts to endow robots with a similar ability have failed to generalize beyond specific tasks, partly because the focus has been on following the trajectory of an action demonstrated by an expert.

The current project investigates a new interdisciplinary approach to imitation learning that is inspired by how humans learn via goal-based imitation. The project's specific objectives include: (1) a new method for imitation based on inferring the underlying goals of human actions rather than following trajectories: actions are executed based on sequences of inferred goals and successfully executed action sequences are cached as higher level goals, leading to hierarchical goal-based imitation; (2) a new approach based on hierarchical Bayesian models (HBMs) is proposed for generalization across objects and tasks, and (3) developmental studies of goal-based imitation learning are proposed for testing predictions of the project?s models in imitation learning experiments with children.

The project represents one of the first efforts to develop rigorous probabilistic models of goal-based imitation learning based on insights from human learning. The results are expected to pave the way for a new generation of machines that can interact fluently with humans, learn new skills from human teachers, and cooperatively solve problems with human partners. The project also provides graduate and undergraduate students with multidisciplinary training in computer science and cognitive science, with K-12 outreach activities aimed at encouraging students from underrepresented groups to pursue careers in science and engineering.

A great deal of research with adults has documented the presence of body maps in the human brain. These neural maps have an organized spatial layout. Neighboring parts of the body are connected in an orderly fashion to areas of the brain that process touch and movement. Body maps are important for many aspect of everyday life including the sense of one's own body and controlling our movements. Body maps also likely play an important role in learning from others, through allowing us to register similarities between ourselves and other people. Despite the importance of body maps, very little is currently understood about how they develop in the early months and years of life. The research supported by this award would provide significant new information on the development of body maps and their relation to early learning. The award supports a collaborative, cross-disciplinary network of investigators who will combine expertise in developmental psychology and infant learning, brain science, cognitive science, computer modelling, and robotics. The proposed network will also support the development and training of junior investigators through specific activities designed to expose them to the benefits of an interdisciplinary approach.

Advances in methods for safely measuring the brain activity of human infants are allowing new questions to be asked concerning the role of body maps in early learning. The proposed research involves using magnetoencephalography (MEG) to non-invasively measure responses of the infant brain to tactile stimulation of different parts of the body (e.g., hands vs. feet), and to relate these responses to aspects of infant learning. Another set of studies involving electroencephalography (EEG) will examine how body maps facilitate early imitation and learning from others. Insights from these studies will inform (and be informed by) a further strain of research using computer modelling that takes bodily factors into account in designing robotic systems that can learn from people. The research questions will also provide insight into the control of brain-computer interfaces that can assist disabled individuals in learning to control artificial limbs and other external devices.

Much knowledge about how human brains process information and generate actions has been informed by carefully controlled experiments in laboratory settings. However, understanding the brain in action requires exploration of its functions outside structured tasks. The current project explores neural processing over many days using large-scale recordings of brain activity augmented with video, audio and depth camera recordings, all simultaneously and continuously monitoring a subject. Importantly, unlike the majority of existing studies, here the subjects receive no instructions but are simply behaving as they wish in their hospital room-including eating, sleeping, and conversing with family. The project will advance data-intensive science and human neuroscience, leveraging external monitoring of the subjects to interpret naturalistic neural activity. The results of this project will be catalytic in understanding of the human brain, opening the door to study of brain function outside the structured confines of laboratory experiments.

The neural decoding algorithms developed will be directly applicable to current Brain-Computer Interfacing (BCI) technologies, enabling the deployment of systems that can predict the user's needs and improve quality of life outside the laboratory. Further, ongoing collaborations with neurosurgeons focus on evaluating this novel data-intensive approach to ethological brain mapping and how it may complement existing clinical functional brain mapping. The project will support and enable the education of students at the intersection of data science and neuroscience, including training scientists at the undergraduate, graduate, and post-doctoral career stages. Results from the research will be distributed as open access publications and code repositories, supporting a commitment to reproducible science.

This proposal focuses on data-driven innovations to enable more accurate decoding and inference of actions from long-term, naturalistic neural recordings. The first aim proposes to develop algorithms for automated decoding of natural motor and speech behaviors. Unsupervised clustering will be used to discover coherent patterns in brain activity, and clusters will be annotated with behaviors automatically parsed from external monitoring streams. Motivated by the size of the dataset and substantial variety between individuals, this scalable computational approach circumvents tedious manual annotation and fine-tuning of parameters. The second aim proposes to infer networks of dynamic causality of cortical networks engaged in task-free, naturalistic behaviors. This aim focuses on testing the hypothesis that neural correlates of naturalistic behaviors differ from those of repeated, instructed behaviors. Functional networks and the dynamic causality of cortical areas will be explored using methods from nonlinear dynamical systems theory. These networks will be compared to results from clinical brain mapping. This project will improve state-of-the-art neural decoding in naturalistic contexts and uncover neural correlates of task-free behaviors in humans.