Search papers, labs, and topics across Lattice.
Australian Institute for Machine Learning, Adelaide University
2
0
5
Residual connections, a staple of modern transformers, actively *harm* quantization robustness by driving activations away from Gaussianity.
Challenging the conventional wisdom that strong certified robustness requires heavy partitioning, this work shows how white-box knowledge of base classifiers in partition-aggregation ensembles can yield significantly tighter robustness guarantees against label-flipping attacks with fewer partitions.
