SUNMONTUEWEDTHUFRISAT
Events for March 21, 2018
-
EE Seminar: IoT in the CMOS Era and Beyond: Leveraging Mixed-Signal Arrays for Ultra-Low-Power Sensing, Computation, and Communication
Wed, Mar 21, 2018 @ 10:30 AM - 11:30 AM
Ming Hsieh Department of Electrical and Computer Engineering
Conferences, Lectures, & Seminars
Speaker: Dr. Siddharth Joshi, University of California, San Diego
Talk Title: IoT in the CMOS Era and Beyond: Leveraging Mixed-Signal Arrays for Ultra-Low-Power Sensing, Computation, and Communication
Abstract: Energy efficiencies obtained by analog processing are critical for next-generation "smart" sensory systems that implement intelligence at the edge. Such systems are widely applicable in areas like biomedical data acquisition, continuous infrastructure monitoring, intelligent sensor networks, and data analytics. However, adaptive analog computing is sensitive to nonlinearities induced by mismatch and noise, which has limited the application of analog signal processing to signal conditioning prior to quantization. This has relegated the bulk of the processing to the digital domain, or a remote server, limiting the system efficiency and autonomy. This talk highlights principled techniques to algorithm-circuit co-design to overcome these obstacles, leading to energy-efficient high-fidelity mixed-signal computation and adaptation.
First, I will provide analytical bounds on the energetic advantages derived by alleviating the need for highly accurate and energy-consuming analog-to-digital conversion through high-resolution analog pre-processing. I will then present an embodiment of this principle in a micropower, multichannel, mixed-signal array processor developed in 65nm CMOS. Spatial filtering with the processor yields 84 dB in analog interference suppression at only 2 pJ energy per mixed-signal operation. At the algorithmic level, I will present work on a gradient-free variation of coordinate descent, Successive Stochastic Approximation (S2A). S2A is resilient to the adverse effects of analog mismatch encountered in compact low-power realizations of high-resolution, high-dimensional mixed-signal processing systems. Over-the-air experiments employing S2A in non-line-of-sight demonstrate adaptive beamforming achieving 65 dB of processing gain.
I will conclude with my vision about the impact of mixed-signal processing on the next generation of computing systems and share my recent work spanning across devices (RRAM), architectures (compute-in memory) and emerging applications(neuromorphic computing). Crossing these hierarchies is critical to leverage emerging technologies in realizing the next generation of sensing, computing, and communicating systems.
Biography: Siddharth Joshi is a Postdoctoral Fellow in the department of Bioengineering at UC San Diego, he completed his PhD in 2017 at the department of Electrical and Computer Engineering, UC San Diego where he also completed his M.S. in 2012. His research focuses on the co-design of custom, non-Boolean and non-von Neumann, hardware and algorithms to enable machine learning and adaptive signal processing in highly resource constrained environments. Before coming to UCSD, he completed a B. Tech from Dhirubhai Ambani Institute of Information and Communication Technology in India.
Host: Alice Parker, parker@usc.edu, x04476
Location: Hughes Aircraft Electrical Engineering Center (EEB) - 248
Audiences: Everyone Is Invited
Contact: Mayumi Thrasher
-
Automated Geometric Shape Deviation Modeling for Cyber-Physical Additive Manufacturing Systems via Bayesian Neural Networks
Wed, Mar 21, 2018 @ 02:00 PM - 03:00 PM
Ming Hsieh Department of Electrical and Computer Engineering
Conferences, Lectures, & Seminars
Speaker: Arman Sabbaghi, Purdue University
Talk Title: Automated Geometric Shape Deviation Modeling for Cyber-Physical Additive Manufacturing Systems via Bayesian Neural Networks
Series: Center for Cyber-Physical Systems and Internet of Things
Abstract: A significant challenge in dimensional accuracy control of a cyber-physical additive manufacturing (AM) system is the comprehensive specification of geometric shape deviation models for different computer-aided design (CAD) inputs on its constituent AM processes. Current deviation model building methods cannot satisfactorily address this challenge in practice because they are unable to leverage previously specified deviation models for different shapes and processes in an automated or rapid manner. We present a new model building methodology based on a class of Bayesian neural networks (NNs) that directly address the challenge of cyber-physical AM systems. Our framework enables automated and computationally efficient deviation modeling of different shapes and/or AM processes without sacrificing predictive accuracy, compared to existing modeling methods on the same samples of manufactured shapes. A fundamental innovation in our framework is the design of new and connectable NN structures that can leverage previously specified models for adaptive and principled model building. The power and broad scope of our method is demonstrated with several case studies on both in-plane and out-of-plane deviations for a wide variety of shapes manufactured under different stereolithography processes. Our Bayesian NN methodology for automated and comprehensive deviation modeling can ultimately be applied to advance fast, flexible, and high-quality manufacturing in a cyber-physical AM system. This talk is based on a paper written by Raquel De Souza Borges Ferreira, Dr. Arman Sabbaghi, and Dr. Qiang Huang.
Biography: Arman Sabbaghi is an Assistant Professor in the Department of Statistics at Purdue University. His research interests include model building for improved control of complex engineering systems, Bayesian data analysis, experimental design, causal inference, and statistical analysis with missing data.
Host: Prof. Paul Bogdan
More Information: sabbaghi-t.jpg
Location: Hughes Aircraft Electrical Engineering Center (EEB) - 132
Audiences: Everyone Is Invited
Contact: Talyia White
