Residue Number System Based Convolution Neural Network Algorithm
Loading...
Date
2022-06-20
Journal Title
Journal ISSN
Volume Title
Publisher
International Journal of Nature And Science Advance Research
Abstract
A number of positively convergent elements
have aided the development of Deep
Learning. The efficiency of floating point
operations is highly optimized in modern microarchitectures.
A whole area of research has
emerged around quantized models, which reduce
by orders of magnitude the amount of required
memory, with a particular focus on quantized
convolution neural networks. However, there is still
a need to rethink how these quantized models can
then be accelerated efficiently. The research starts
by recognizing that inference in convolution neural
networks is fundamentally a Digital Signal
Processing (DSP) task. Memory, computation, and
power utilization were expensive, in a different but
similar way to how they are on modern mobile
platforms, whose budget is set by their battery
capacity. It is therefore of
utmost importance to
provide an alternative
solution to this problem.
This research introduces
Residue Number System
Architecture to the
process to take advantage
of the limited - but not
binary - range of values
that the operands can
assume during the
convolution operation in a