Jeremy Kun
Jeremy Kun
An example of generating a well-fitting polynomial using lolremez: https://github.com/samhocevar/lolremez/issues/28#issuecomment-1913324892 Another tool: https://github.com/pychebfun/pychebfun
Outline sketch: - Use an existing Remez solver to find any old approximation to sgn or max of some arbitrary degree, e.g., from https://github.com/google/heir/issues/658#issuecomment-2087937786 - Implement a lowering to that...
Thanks to Seonhong Min for sending me https://eprint.iacr.org/2018/462, in which it shows that BFV achieves polynomial approximations via a fixed-point approximation, not a floating point one. I think there is...
After starting an implementation in https://github.com/google/heir/pull/665 (with a fixed approximation polynomial) and discussing in the HEIR meeting today, I have a few kinks to work out: 1. I want to...
These folks do something slightly different, which is more holistic in re NN training: https://github.com/EfficientFHE/SmartPAF They pick small degree polynomial approximations and then do a variety of network fine-tuning to...
I added an upstream RFC for the polynomial approximation pass https://discourse.llvm.org/t/rfc-a-polynomial-approximation-pass/79301
That's a good point, I don't see how it could be related to anything except LWE/RLWE.
Part of https://github.com/google/heir/issues/475
Another example specific to OpenFHE, we hard-code `Ciphertext` regardless of what encoding the mlir specifies.
I think in the very long term we would bypass the libraries. In the medium term we could work with the OpenFHE developers to support overriding their calculations (perhaps by...