Hyperlyve: Parallelized Enumeration and Exact Lyapunov Stability Certification for Deep ReLU Networks
The authors report enumerating ReLU networks over 150 times faster than competing tools, enabling stability proofs that current verifiers cannot handle.
Hyperlyve learns candidate Lyapunov functions and formally checks them in a loop guided by counterexamples. Its parallel algorithm enumerates the regions formed by deep networks’ ReLU activations and checks stability conditions in each. The tests certify the exact smooth dynamics within a compact domain, rather than a piecewise-linear approximation. The paper also establishes that verifying these continuous piecewise-affine neural Lyapunov functions is co-NP-complete.
ArXiv · AI/CL/LG's note
Hyperlyve learns candidate Lyapunov functions and formally checks them in a loop guided by counterexamples. Its parallel algorithm enumerates the regions formed by deep networks’ ReLU activations and checks stability conditions in each. The tests certify the exact smooth dynamics within a compact domain, rather than a piecewise-linear approximation. The paper also establishes that verifying these continuous piecewise-affine neural Lyapunov functions is co-NP-complete.
ArXiv · AI/CL/LG's note
score 4