The problem ZeroGravity solves

While significant progress has recently been made to create zero-knowledge proofs of training and inference of neural networks, the algorithms and data types of popular neural network architectures fundamentally don't lend themselves well to being expressed as arithmetic circuits.

Weightless Neural Networks (WNNs) present an alternative, using table lookups instead of floating point arithmetic to perform inference. We exploit recent acheivements in lookup arguments to reduce the proving of inference to lookup arguments on a highly optimised Bloom filter implementation.

As a concrete use case, this allows a person to prove that their biometric data (derived from an ML model) is not on a blacklist, without having to reveal their private data, allowing for proving data provenance in ML-based medical scans.

Challenges we ran into

https://hackmd.io/@benjaminwilson/zero-gravity#A-challenge-overcome-the-choice-of-hash-function