Science and Engineering Facility Room 213
1414 Circle Drive
University of Tennessee
Knoxville, TN 37996-2100 Skype
Research
Fabric on a Chip (FoC)
Hardware Recurrent Neural Networks Architectures
I worked on resource-efficient realizations of neural networks in hardware. The approach I took is that of simplifying the arithmetic involved by using piece-wise linear approximations for all operations. At a modest reduction in accuracy this methodology offers substantial gains in area and speed.
Selected Publications
[1]
V.
Mahoney,
I.
Elhanany,
"A Backpropagation Neural Network Design Using Adder-Only Arithmetic,"
in Proc. 2008 IEEE Midwest Symposium on Circuits and Systems (MWSCAS),
August,
2008
[pdf]