S4 achieved state of the art results in many long sequence benchmarks
(Treating mnist as a sequence and finish sequence after 400 pixels using s4)
0 Epoch / 100 Epoch / 200 Epoch
The Annotated S4: https://srush.github.io/annotated-s4/
The Annotated S4D: https://srush.github.io/annotated-s4/s4d.html
https://github.com/HazyResearch/state-spaces
HiPPO: https://arxiv.org/pdf/2008.07669.pdf
Linear State Space Layer: https://arxiv.org/pdf/2110.13985.pdf
Structured State Spaces: https://arxiv.org/pdf/2111.00396.pdf
Diagonal Structured State Space: https://arxiv.org/pdf/2203.14343.pdf
HiPPO blog: https://hazyresearch.stanford.edu/blog/2020-12-05-hippo
Presentation from author on s4: https://www.youtube.com/watch?v=luCBXCErkCs