AI Verification and Interpretation
Verification across neural architectures and analysis of complex AI reasoning.
Trustworthy AI requires methods matched to different architectures, scales, and applications. This project combines formal verification for models that admit explicit analysis with interpretation and monitoring for systems whose reasoning is difficult to verify end to end.
Neural-network verification
NNV3 expands neural-network verification to additional architectures and application domains. Our work on first-to-fire spiking neural networks studies the robustness of spike-based decisions and remains under review at TACAS 2027.
The SNNV interactive site explains the temporal verification problem, the set representation used for first-to-fire networks, and the associated examples.
Neuro-symbolic reasoning
The SAIV 2026 work A Self-correcting Neuro-symbolic AI Reasoning Framework combines neural reasoning with symbolic checks that identify invalid intermediate steps and support correction. Related work studies probabilistic verification across vision, speech, and depth models.
Interpreting AI reasoning
For larger language and reasoning models, my ongoing collaborations study how to locate errors, characterize model behavior, and guide corrective intervention. This page reports that research direction without importing unsupported numerical claims from exploratory project pages.