Apple proposes Self-Proving models that generate correct outputs and prove their correctness to a verification algorithm V via interactive proofs. This addresses the gap in traditional accuracy metrics, which only average over distributions without per-input guarantees. Models are trained to succeed with high probability on sampled inputs from a given distribution.
Key Points
- 1.Proposes training models to generate correct outputs and interactive proofs
- 2.Provides per-input correctness guarantees beyond average accuracy
- 3.Uses verification algorithm V for proof validation
- 4.High-probability success over input distributions
Impact Analysis
Enables trustworthy AI for safety-critical apps by proving specific predictions. Boosts adoption in regulated industries needing verifiability. Shifts focus from statistical to provable correctness.
Technical Details
Models interact with verifier V in an interactive proof protocol. Training ensures soundness: correct output implies valid proof, and completeness: correct models generate provable outputs. Theoretically founded on probabilistic guarantees.