Closed jamesjingqili closed 2 months ago
KFSB is just a small variation of FSB. It picks the top-k neurons among all layers to evaluate, whereas the original FSB score picks the top-k neuron for each layer, and thus can be a bit slow when the network has many layers. For efficiency, we use KFSB by default. If I recall correctly, this heuristic was introduced in our VNN-COMP 2021 submission to address the high overhead of branching heuristics. You can still cite the original FSB paper and mention that α,β-CROWN has a variation of FSB to improve its efficiency on deeper networks.
Thank you very much for your timely reply!
If we want to cite the KFSB method, then which paper should we cite? I tried to search a lot, but I can only find the paper that proposed FSB. Any pointer will be appreciated. Thanks!