Project data description:
Using a tool called condensa I can prune values from a DNN model and
retrain to recover accuracy resulting in a sparsely populated set of
weights achieving similar accuracy. The resulting sparse model contains
significantly less (<1%) data, which presents an opportunity for a sparse
implementation which only computes the result for nonzero elements of the
model. Unfortunately, such an implementation fails to leverage modern
devices, which have been tailored to efficiently execute large, dense,
highly data parallel operations. The long term vision of my project is to
analyze the pros and cons of sparse implementations of such operations, and
identify opportunities to exploit locally dense clusters of data in the
model.

Viz Goal and target audience:
I will be visualizing sparse weights in a single convolution layer of a
DNN. The visualization will help guide the sparse implementation of the
convolution operation, and will eventually accompany the performance
statistics, as well as architectural implications of the implementation on
various hardware. The data will be a sparse 4-d matrix. Since I am
interested in the data at the filter level, it will be represented as a 2d
matrix of 3x3 filters. I will be visualizing local sparsity (at the filter
level) as well as global sparsity. Possible other areas of interest may be
visualizing the filters as vectors alongside density information to inform
a cache prefetcher, or to visualize the density implications of altering
the pruning value. The target audience will be the reading committee of my
masters project, who will be professors with a background in computer
architecture and an insight into performance critical, concurrent software.