Fast SDP certification of neural networks : towards large multi-class datasets

What the math gives to ML

The paper's transferable contribution is a verifier formulation that treats the worst target class as an internal optimization variable, rather than solving one robustness problem per incorrect class. This converts multi-class certification into a shared-margin problem and can reduce repeated propagation, lifted variables, and SDP solves when the number of classes is large. The second useful asset is structure-aware neuron pruning: neurons whose activation sign or bounded contribution is irrelevant to the worst-case margin can be removed from the relaxation while preserving a sound certificate. These ideas are most promising in certified training and large-class robustness evaluation, where verifier cost rather than forward inference is the bottleneck.