GPU Kernels for Block-Sparse Weights

This work releases highly optimized GPU kernels for an underexplored class of neural network architectures: networks with block-sparse weights, which improve upon the state-of-the-art in text sentiment analysis and generative modeling of text and images.