Event Details
A Modified Minimal Gated Unit and Its FPGA Implementation
Presenter: Tong Zhu
Supervisor:
Date: Thu, December 10, 2020
Time: 14:00:00 - 15:00:00
Place: ZOOM - Please see below.
ABSTRACT
Join Zoom Meeting:
https://uvic.zoom.us/j/89751224035?pwd=c1RPaGpXbFF5cGxtaG54THBJS21KUT09
Meeting ID: 897 5122 4035
Password: 006031
(Note: you can only join via Zoom SSO)
Abstract:
Recurrent neural networks (RNNs) are versatile structures used in a variety of sequence data-related applications. The two most popular proposals are long short-term memory (LSTM) and gated recurrent unit (GRU) networks. Towards the goal of building a simpler and more efficient network, minimal gated unit (MGU) has appeared and shown quite promising results. In this project, we present a simple and improved MGU model, MGU_1, implemented on scalable field programmable gate arrays (FPGA). Experiments with various sequence data show that the MGU_1 has better accuracy compared to the MGU. The accelerator implemented on FPGA accelerates the inference phase utilizing the model trained on our indoor localization data set. It has two layers of MGU_1 and each has 32 hidden units. The accelerator can achieve 142 MHz and 60 GOPS on the Xilinx XC7Z020 FPGA and outperforms the Intel i5-5350U based software solution by two orders of magnitude.