Accelerating Sparse Neural Networks on GPUs

Alexander Ertl

Supervisor(s): Markus Steinberger

TU Graz

Abstract: Ever larger networks with parameters in the order of the hundreds of millions are required to fit increasingly complex and expansive datasets. In conjunction with ubiquitous machine learning applications on mobile or embedded platforms, this makes efficiency a vital property of artificial neural networks. Therefore we build upon work on replacing fully connected dense layers with trainable, evolving sparse layers in CSR encoding. This allows us to train networks at sparsity levels of up to 97% while considerably reducing the memory footprint as well as the number of computations thereby indicating that GPU accelerated sparse layers are a viable alternative to dense layers.
Keywords: Computer Vision, Graphics Hardware
Full text:
Year: 2021