Publication
Systems Group Master's Thesis, no. 160; Department of Computer Science, March 2017
Supervised by: Prof. Ce Zhang
Supervised by: Prof. Ce Zhang
We rst present an application where we helped our astrophysicist collaborators
to recover features from arti cially degraded images with worse
seeing and higher noise than the original with a performance which far exceeds
simple deconvolution by training a generative adversarial network
(GAN) on galaxy images. However, training time is limiting our potential
to train on larger data sets. It takes 2 hours to train a GAN using
4105 galaxy images for 20 iterations on an NVIDIA TITAN X GPU. We
ask the question: Can we speed up our machine learning training process
by reducing the precision of data representation?...
@mastersthesis{abc,
abstract = {We rst present an application where we helped our astrophysicist collaborators
to recover features from arti cially degraded images with worse
seeing and higher noise than the original with a performance which far exceeds
simple deconvolution by training a generative adversarial network
(GAN) on galaxy images. However, training time is limiting our potential
to train on larger data sets. It takes 2 hours to train a GAN using
4105 galaxy images for 20 iterations on an NVIDIA TITAN X GPU. We
ask the question: Can we speed up our machine learning training process
by reducing the precision of data representation?...},
author = {Hantian Zhang },
school = {160},
title = {The ZipML Framework for Training Models with End-to-End Low Precision},
year = {2017}
}