Search

Your search keyword '"Rekhi, Angad S."' showing total 7 results
Start Over Author "Rekhi, Angad S."
7 results on '"Rekhi, Angad S."'

1. Remote Sub-Wavelength Focusing of Ultrasonically Activated Lorentz Current

Author

Rekhi, Angad S. and Arbabian, Amin

Subjects
Physics - Medical Physics
Abstract

We propose the use of a combination of ultrasonic and magnetic fields in conductive media for the creation of RF electrical current via the Lorentz force, in order to achieve current generation with extreme sub-wavelength resolution at large depth. We demonstrate the modeling, generation, and measurement of Lorentz current in conductive solution, and show that this current can be localized at a distance of 13 $cm$ from the ultrasonic source to a region about three orders of magnitude smaller than the corresponding wavelength of electromagnetic waves at the same operation frequency. Our results exhibit greater depth, tighter localization, and closer agreement with prediction than previous work on the measurement of Lorentz current in a solution of homogeneous conductivity. The proposed method of RF current excitation overcomes the trade-off between focusing and propagation that is fundamental in the use of RF electromagnetic excitation alone, and has the potential to improve localization and depth of operation for RF current-based biomedical applications., Comment: 5 pages, 7 figures; submitted to Applied Physics Letters

Published
2017
Full Text
View/download PDF

5. Analog/Mixed-Signal Hardware Error Modeling for Deep Learning Inference.

Author

Rekhi, Angad S., Zimmer, Brian, Nedovic, Nikola, Ningxi Liu, Venkatesan, Rangharajan, Miaorong Wang, Khailany, Brucek, Dally, William J., and Gray, C. Thomas

Subjects
DEEP learning, MIXED signal circuits, ENERGY consumption, SYSTEMS design, ANALOG multipliers
Abstract

Analog/mixed-signal (AMS) computation can be more energy efficient than digital approaches for deep learning inference, but incurs an accuracy penalty from precision loss. Prior AMS approaches focus on small networks/datasets, which can maintain accuracy even with 2b precision. We analyze applicability of AMS approaches to larger networks by proposing a generic AMS error model, implementing it in an existing training framework, and investigating its effect on ImageNet classification with ResNet-50. We demonstrate significant accuracy recovery by exposing the network to AMS error during retraining, and we show that batch normalization layers are responsible for this accuracy recovery. We also introduce an energy model to predict the requirements of high-accuracy AMS hardware running large networks and use it to show that for ADC-dominated designs, there is a direct tradeoff between energy efficiency and network accuracy. Our model predicts that achieving <0.4% accuracy loss on ResNet-50 with AMS hardware requires a computation energy of at least ~300 fJ/MAC. Finally, we propose methods for improving the energy-accuracy tradeoff. [ABSTRACT FROM AUTHOR]

Published
2019
Full Text
View/download PDF

Catalog

Books, media, physical & digital resources
See catalog results