Abstract Neural Networks
summary
Paper claims a procedure to compress a large DNN into a smaller Abstract Neural Network (ANN) that soundly over-approximates every input-output behavior, and gives a constructive algorithm to build it. The algorithm merges nodes, forms layer abstracts by taking the weighted convex hull (extreme points suffice for convex domains), then proves soundness under two conditions: non-negative post-activation intermediates and the Weakened Intermediate Value Property for activations. Code and proofs are provided and examples show where the conditions fail and how to fix them.
AI-written summary · about
chapters (4)
- 0:00Start
- 3:22Contributions
- 3:37Interval Neural Networks (INN)
- 18:16Defined Abstract Neural Networks (ANNS)
related talks
31mAbstract Neural Networks
56mAbstract Domains for Machine Learning Verification
19mInterval Universal Approximation for Neural Networks
11mAGNES: Abstraction-guided Framework for Deep Neural Networks Security
58mBuilding Trust and Safety in Artificial Intelligence with Abstract Interpretation
31mARENA: Enhancing Abstract Refinement for Neural Network Verification
01h 51mUnderstanding Neural Networks (Somewhat)
18mTaming Reachability Analysis of DNN-Controlled Systems via Abstraction-Based Tr...