Certified Circuits: Provable Stability for Neural Network Interpretability
The newly introduced framework, Certified Circuits, offers verifiable stability assurances for the mechanistic interpretability of neural networks. Current methods for circuit discovery are fragile, as they rely heavily on the selected concept dataset and frequently struggle to generalize beyond their training data. Certified Circuits enhances any black-box discovery technique by employing randomized data subsampling, ensuring that the decisions regarding circuit elements—such as neurons or edges in the model graph—remain consistent despite minor alterations to the concept dataset. This approach avoids unstable components, resulting in more dependable circuits. The findings are detailed in a publication on arXiv (2602.22968) and seek to enhance the debugging, auditing, and deployment processes of neural networks.
Key facts
- Certified Circuits provides provable stability guarantees for circuit discovery.
- Existing circuit discovery methods are brittle and dataset-dependent.
- The framework uses randomized data subsampling.
- It certifies invariance to bounded edit-distance perturbations.
- Unstable components are abstained from.
- The research is published on arXiv with ID 2602.22968.
- Mechanistic interpretability aims to identify minimal subnetworks responsible for specific behaviors.
- The work addresses out-of-distribution transfer failures.
Entities
Institutions
- arXiv