Most of the research in verifiable ML is on Zero-Knowledge Machine Learning (zkml). However, zkml currently has a performance overhead of 1000x and cannot run large models yet. Source: Mohamed Baioumy & Alex Cheema from AI x Crypto Primer

Several approaches are being explored. This is a relatively new field and there is an opportunity for different approaches that make different tradeoffs. Ora is experimenting with opML. This approach involves a single party ‘optimistically’ inferencing a model, putting the result on-chain, and incentivising verifiers to challenge incorrect results by paying them tokens. Aizel is building a solution based on Multi-Party-Computation (MPC) and Trusted Execution Environments (TEE). Their aim is to do verifiable inference at the same cost as normal inference.

Mohamed Baioumy & Alex Cheema from AI x Crypto Primer - full credits go toward these legends.