Verified Deep Learning with Lean 4
Formal Backpropagation from MLP to Attention, via MLIR

Acknowledgments

Thanks to Tomáš Skřivan, whose Scientific Computing in Lean bootstrapped this project’s first MNIST trainer.

Thanks to Monica Omar on Lean Zulip, who closed the 6-level Finset.sum commutation in the Conv2D VJP proof.

Thanks to Franz Huschenbeth on Lean Zulip for reviewing early drafts and suggesting the comparator approach used in Appendix B.