Details
Presenter(s)
Display Name
Jacob Nelson
- Affiliation
-
AffiliationTennessee Technological University
- Country
Abstract
WORDA is a convolution algorithm developed with the aim of reducing the latency of convolutional neural networks deployed on FPGA. The Winograd convolution algorithm is utilized and expanded on through the use of a decomposition scheme as well as a different set of transformation matrices compared to other implementations. This greatly reduces the latency of the algorithm while including the flexibility to compute convolutions with a filter larger than 3x3.