2 comments

[ 2.9 ms ] story [ 16.1 ms ] thread
I recently made a similar implementation for SHA256. Unfortunately it suffers from gradient vanishing. Source code: https://github.com/calebh/differentiablesha256
Cool project! Yes, it would be really surprising if this "worked," because SHA256 is meant to be very difficult to reverse. But it's interesting that the failure mode is gradient vanishing as opposed to, say, getting trapped in a local optimum.