The goal of DoubleBlockSparse project is to implement custom layers for popular
machine learning frameworks that enable double sparsity eg. both sparse
weights and sparse features. It combines ideas from several references:
block sparsity from [1], double sparsity from [2], local, biologicaly
plausable sparse learning rules based on lateral inhibition [3], and
sparse evolutionary training [4].

[1] - https://openai.com/blog/block-sparse-gpu-kernels

[2] - https://numenta.com/neuroscience-research/research-publications/posters/icml-2019-how-can-we-be-so-dense

[3] - https://www.ibm.com/blogs/research/2019/04/biological-algorithm

[4] - Decebal Constantin Mocanu, Elena Mocanu, Peter Stone, Phuong H. Nguyen, Madeleine Gibescu, Antonio Liotta - Scalable Training of Artificial Neural Networks with Adaptive Sparse Connectivity inspired by Network Science - https://arxiv.org/abs/1707.04780