UNIVERSITY PARK, Pa. — Abhronil Sengupta, the Joseph R. and Janice M. Monkowski Career Development Assistant Professor of Electrical Engineering at Penn State, was granted a three-year, $360,000 Early Career Program Award from the Army Research Office (ARO). The award supports “early career scientists and engineers who show exceptional ability and promise for conducting basic research,” according to the Army Research Laboratory (ARL).
The ARO Early Career Program Award targets scientists and engineers who have held a tenure track position at a U.S. institution of higher education for fewer than five years at the time of application. According to an announcement published by the ARL, the award’s objective is “to foster creative basic research in science and engineering; enhance development of outstanding early career investigators; and increase opportunities for early career investigators to pursue research in areas relevant to the Army.”
Sengupta leads the Neuromorphic Computing Lab at Penn State, where his group explores next-generation, brain-inspired artificial intelligence systems that forge stronger connections with neuroscience in order to circumvent algorithmic and hardware scaling challenges of current deep learning solutions. His work on neuromorphic computing has also been recognized with a NSF CAREER Award, IEEE Electron Devices Society Early Career Award and the IEEE Circuits and Systems Society Outstanding Young Author Award, among others.
Penn State News spoke with Sengupta about the research to be conducted with the grant.
Q: What is the goal of this research?
Sengupta: Current brain-inspired engineered systems have primarily focused on the emulation of bio-plausible computational models of neurons and synapses, but incorporation of other cellular units from the brain is lacking. The Early Career project explores a holistic system-science enabled perspective to design brain-inspired learning systems by understanding the role of astrocytes. Astrocytes are an under-explored yet critical component of the brain responsible for enabling rich temporal dynamics such as neural synchronization that form the dynamical basis for learning and memory. Specifically, our project focuses on spinal central pattern generators (CPGs). CPGs are neural circuits that generate spontaneous rhythmic patterns and serve as a “brain-like” model system to design engineered learning platforms. Building upon theoretical neuroscience insights, the project will address the unmet need of understanding the key aspects of bio-fidelity required for the design of astromorphic algorithms and hardware in the context of autonomous robotic locomotion tasks.
Q: How do you plan to achieve these goals?
Sengupta: The research involves a transformative research agenda at the intersection of theoretical neuroscience, algorithms and hardware to decode and embed astrocyte functionality in brain-inspired algorithm and hardware design. The project spans complementary and intertwined explorations across multiple focus areas, ranging from neuromorphic CPG modelling, algorithmic formulations to design CPGs for robotic locomotion tasks capable of online local learning and novel energy-efficient, spin-based devices and circuits that inherently exploit the intrinsic device physics for mimicking astrocyte functionalities.
Such an end-to-end framework combining principles across the life, physical and computer sciences has the potential to enable a paradigm shift in the design of dynamical intelligent systems and have a long-term impact on the National Academy of Engineering’s Grand Challenge related to “Reverse-Engineer the Brain." Implementation of energy-efficient astromorphic hardware and algorithms will be critical for a large spectrum of Department of Defense applications for enabling real-time intelligence in autonomous systems like unmanned military robotic systems, among others.
Q: How does your research differ from what’s already been done regarding robotic locomotion?
Sengupta: In recent years, global policy optimization for robotic locomotion through reinforcement learning has been widely investigated. However, training such systems is computationally challenging, limiting its applicability for on-chip learning in edge robotic systems with resource constraints.
Existing work on brain-inspired CPG uses over-simplified neuron models and network architectures. The lack of inclusion of biological details in the CPG architecture has resulted in limited flexibility of the control system, thereby constraining their applicability to mostly simple robotic platforms like hexapod robots in contrast to the more complex design space of real-world robotic locomotion control of more intensively researched models, such as quadruped robots. In this project, we develop a detailed bio-inspired CPG model and propose to show that astrocyte control is instrumental to ensure optimal and stable gait emergence in robotic quadruped locomotion systems. Local learning mediated gait emergence ensures compatibility of our proposed control system with neuromorphic hardware, thereby leading to the potential of enabling real-time, low-power on-chip learning.
Q: In what ways has Penn State helped to enable this research?
Sengupta: The focus on multidisciplinary research at Penn State and our lab’s affiliation and collaboration with researchers in the Center for Artificial Intelligence Foundations and Engineered Systems (CAFE) and Materials Research Institute (MRI) at Penn State has been instrumental in enabling this research. I am also highly indebted to the U.S. National Science Foundation which funded our initial research in the field of astrocyte based neuromorphic computing through the Early Concept Grant for Exploratory Research (EAGER) program which is specifically targeted for interdisciplinary high-risk, high-payoff projects with a transformative scope.
Q: What are you most excited about?
Sengupta: The cross-cutting nature of the project with multiple disciplines excites me the most. The key distinguishing point of our project lies in the fact that we are striving to utilize computational neuroscience insights of astrocyte signaling enabled CPG controller design for legged robotic locomotion while parallelly investigating novel devices and circuits which can mimic astrocyte functionalities through their intrinsic physics. We believe such an interplay across the stack of theoretical neuroscience, algorithms and hardware can result in the development of a new generation of computationally efficient autonomous robotic neuromorphic platforms that are able to adapt to changing environment in an online fashion.