Linear Bound Propagation
A classical class of network verification methods that processes a network operation by operation to propagate linear bounds. At second-order terms it pays the sum of the per-operation maxima rather than the coupled maximum, which results in a loss of tightness in the derived bounds.
Timeline (last 90 days)
GEAR is a GPU-accelerated global nonlinear programming (NLP) solver whose core idea originates from the linear bound propagation framework in neural network verification.
GEAR uses linear bound propagation as a relaxation means for nonlinear constraints, provides lower bounds for branch and bound, and performs safe pruning when a subregion is detected to be infeasible.
Linear bound propagation is a core technique in the formal verification of neural networks, used to compute safety bounds on neural network outputs with respect to input perturbations.
The linear bound propagation framework has been widely applied in verification tools such as ERAN and CROWN.
On GPT-2, the moment relaxation certificate is tighter than linear bound propagation on every attention head.
The looseness of linear bound propagation exactly matches the theoretically predicted per-operation excess.
Per-operation relaxation (including linear bound propagation) computes, at the second-order term, the sum of the maxima of each operation rather than a coupled maximum, resulting in a loss of tightness.
All Facts (7)
GEAR is a GPU-accelerated global nonlinear programming (NLP) solver whose core idea originates from the linear bound propagation framework in neural network verification.
50%UnverifiedGEAR uses linear bound propagation as a relaxation means for nonlinear constraints, provides lower bounds for branch and bound, and performs safe pruning when a subregion is detected to be infeasible.
50%UnverifiedLinear bound propagation is a core technique in the formal verification of neural networks, used to compute safety bounds on neural network outputs with respect to input perturbations.
50%UnverifiedThe linear bound propagation framework has been widely applied in verification tools such as ERAN and CROWN.
50%UnverifiedOn GPT-2, the moment relaxation certificate is tighter than linear bound propagation on every attention head.
50%UnverifiedThe looseness of linear bound propagation exactly matches the theoretically predicted per-operation excess.
50%UnverifiedPer-operation relaxation (including linear bound propagation) computes, at the second-order term, the sum of the maxima of each operation rather than a coupled maximum, resulting in a loss of tightness.
50%