What Should We Do If the Bound Is Inconclusive? Neural Network Certification, Part 5
Split April’s input region, bound each piece, and assemble the local results into a complete certificate.
Recent posts
Why Do Simple Bounds Lose Information? Neural Network Certification, Part 4
Triangle recovers relationships that intervals forgot. DeepPoly carries selected linear bounds, while CROWN-IBP trades some information for speed.
How Can Bounds Travel Through a Network? Neural Network Certification, Part 3
Follow ranges through April’s tiny classifier and turn a lower bound on its output margin into a certificate.
What Would Count as a Proof? Neural Network Certification, Part 2
April meets a tiny ReLU classifier as we turn an informal promise into a mathematical statement and prove it by hand.
Why Testing Is Not Proof: Neural Network Certification
April the Siberian cat helps us see why an attack can disprove robustness, while only a certificate can prove it.