DARPA Young Faculty Award to Develop Ultra-Low Power Machine-Learning Hardware
ECE Associate Professor Aatmesh Shrivastava was awarded up to $1M Young Faculty Award from DARPA for “Nano-Watt Power Machine-Learning Hardware Using Precision Analog Computing.”
Aatmesh Shrivastava, associate professor of electrical and computer engineering, is developing very low power, analog-based machine learning (ML) hardware that will result in sophisticated vision applications that would not be feasible with digital technology. Shrivastava’s research earned him a highly prestigious DARPA Young Faculty Award for up to $1 million.
His team will focus on creating an integrated system-on-a-chip (SoC), which will in initial phases recognize simple images, such as animals and characters.
The goal is to scale up the technology to identify and categorize sophisticated imaging. While the research will initially apply to military applications, it could eventually have an impact on a variety of industries. Shrivastava sees it playing a role in autonomous driving, target recognition, and machine vision for drones, for starters.
“It can have a huge, broad impact because it is basically trying to improve upon the fundamental way of computing that has been established for several decades,” Shrivastava says of the significance of choosing analog over digital technology as the basis for this research.
A digital computing device could not match the efficiencies of analog and would instead hit a wall because of limitations around processing, size, and power consumption. “Inherently, its process is inefficient when it comes to utilizing hardware,” Shrivastava adds.
The project components will include ML hardware with robust and precise analog computing circuits, an analog computing system modeling tool, and an adversarial attack detector.
Shrivastava has a substantial body of work related to analog technology as well as power consumption and delivery. He is the recipient of a prestigious NSF CAREER award for “An Ultra-low Power Analog Computing Hardware Design Framework for Machine Learning Inference in Edge Biomedical Devices.” This work is focusing on developing ML-based hardware solutions for wearable and implantable biomedical devices using ultra-low power analog computing circuits.
In collaboration with Cornell University, Shrivastava and ECE Professor Nian Sun, received a $2 million NSF grant for “Heterogeneous Integration in Power Electronics for High-Performance Computing (HIPE-HPC).”
Additionally, he was awarded a patent for “Self-powered analog computing architecture with energy monitoring to enable machine-learning vision at the edge.”
Shrivastava sees the awards and funding as recognition of analog computing and “a real signal of its importance.”
Concurrent with this research, Shrivastava has launched Think Analog, a startup company that he hopes will one day produce commercialized products based on this low-power, analog computing framework.
Technical Summary: ECE Associate Professor Aatmesh Shrivastava was awarded up to $1M Young Faculty Award (YFA) from DARPA for “Nano-Watt Power Machine-Learning Hardware using Precision Analog Computing.” This YFA project aims to realize ultra-low power (nano-watt level), analog computing, machine-learning (ML) hardware for applications at the edge that are otherwise not possible due to power consumption. The project will also involve the development of a detection technique to protect against easily implemented adversarial attacks. The machine-learning (ML) hardware platform will include the development of robust and precise analog computing circuits, an analog computing system modeling tool, a current sensing-based adversarial attack detector, and an integrated system-on-a-chip (SoC) design for ML vision application for demonstration. The project aims to overcome the longstanding barriers to reduce the device’s power consumption and size, particularly for edge applications. It will include a demonstration of multi-layer computing previously not demonstrated in analog. The SoC will aid DoD applications such as autonomous driving, target recognition, machine vision for drones, among others.