Jeng-Hau Lin
Jeng-Hau Lin has a total of 4 co-authors. Here are the top co-authors by number of shared publications:
Mulong Luo
School of Electrical and Computer Engineering, Cornell University
1 shared publicationRajesh K. Gupta
Dept. of Computer Science and Engineering
1 shared publication- Xun Jiao
Villanova University, Villanova, PA, USA
1 shared publication - Zhuowen Tu
Department of Cognitive Science, University of California, San Diego, CA, USA
1 shared publication
Latest Publications
- 9 April 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Design & Test
- Vol. 37 (5) , 75-84
Vulnerability of Hardware Neural Networks to Dynamic Operation Point Variations
Deep Neural Networks have found many practical applications that are beginning to demand their implementation in hardware for a variety of cost, performance reasons. Microelectronic variability, that is, variations in manufactured hardware to variations in operating conditions are a practical reality that has led to design guardbands. These guardbands increasingly undermine the cost/performance advantage of hardware implementations and force us to seek more efficient solutions often spanning the entire system design. In this paper, we explore to what extent microelectronic variations affect the quality of results from neural networks and then understand with what opportunities we can improve networks to survive the challenges aroused by physical variations. Given the inherent resilience of neural networks due to the adaptation of their learning parameters, one would expect the quality of results produced by neural networks to be relatively insensitive to the rising timing error rates caused by increased variations. On the contrary, our results of error injection on multiple-layer perceptron (MLP) and convolutional neural network (CNN) show that the two canonical network types cannot withstand typical physical variations with voltage droops as small as 20 mV. Instead, alternative algorithmic implementations provide greater resilience. In particular, we examine here two such alternatives: the binarized neural network (BNN) and local binary pattern network (LBPNet) and present results that show greater sustenance against physical variations.
- 9 April 2020
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Design & Test
- Vol. 37 (5) , 75-84
Vulnerability of Hardware Neural Networks to Dynamic Operation Point Variations
Deep Neural Networks have found many practical applications that are beginning to demand their implementation in hardware for a variety of cost, performance reasons. Microelectronic variability, that is, variations in manufactured hardware to variations in operating conditions are a practical reality that has led to design guardbands. These guardbands increasingly undermine the cost/performance advantage of hardware implementations and force us to seek more efficient solutions often spanning the entire system design. In this paper, we explore to what extent microelectronic variations affect the quality of results from neural networks and then understand with what opportunities we can improve networks to survive the challenges aroused by physical variations. Given the inherent resilience of neural networks due to the adaptation of their learning parameters, one would expect the quality of results produced by neural networks to be relatively insensitive to the rising timing error rates caused by increased variations. On the contrary, our results of error injection on multiple-layer perceptron (MLP) and convolutional neural network (CNN) show that the two canonical network types cannot withstand typical physical variations with voltage droops as small as 20 mV. Instead, alternative algorithmic implementations provide greater resilience. In particular, we examine here two such alternatives: the binarized neural network (BNN) and local binary pattern network (LBPNet) and present results that show greater sustenance against physical variations.