Boosting Multi-Neuron Convex Relaxation for Neural Network Verification
summary
A volume-approximation pipeline selects which neuron groups get true convex-hull relaxation instead of computing all 2K-dimensional hulls: FaGMR computes over/under-approximated volumes from K-dimensional relaxations, picks groups with small overestimates and drops groups with large underestimates, then runs the SPM convex-hull generator only on the selected sets. Evaluation on MNIST and CIFAR-10 with K=3 shows FaGMR reduces group count and LP solving time, avoids many OOM cases, and is faster with equal or better precision than prior state-of-the-art. K-dimensional substitute for 2K computations
AI-written summary · about
chapters (10)
- 0:00Neural network verification intro
- 2:17Why verification is needed
- 4:10How DNNs process inputs
- 6:22Limitations of single-neuron relaxations
- 8:54Multi-neuron relaxations intro
- 10:23Static group selection problems
- 13:35Dynamic selection with Figma
- 19:06MNIST and CIFAR experiments
- 24:54Figma advantages summary
- 25:59Question and answer session
related talks
23mPRIMA: General and Precise Neural Network Certification via Scalable Convex Hull
14mConvex Hull Approximation for Activation Functions
20mIncremental Verification of Neural Networks
42mTowards Verification of General Neural Networks: A Dual Approach
20mSupport Triangle Machine
25mProof transfer for fast certification of multiple approximate neural networks
46mSmooth Nonconvex Min-Max Optimization
43mComputational and Statistical Tradeoffs via Convex Relaxation