2a7157c84dcf263f77b37d6c11d7d149-Supplemental-Conference.pdf
–Neural Information Processing Systems
Onecanobservethatf(x)= 1 ( 1) log ( ( 1)x+ 1)isof 1-Lipschitzfunctionforany 6=0,1. We batch-normalization exceptfor 2-layer MLP connected similarity: f(x, y; )= x>y/( kxk2kyk2), with =0.5.
Neural Information Processing Systems
Feb-8-2026, 01:35:13 GMT