Inquiry on results of ConvBig-cifar10 with $r_u$.

[Dalnaracrater]

Hello,
Thank you for your intriguing work to improve verified accuracy and verification time.

I have tried to run your work on Conv-big (Cifar 10) with $r_u$ to understand it better but have encountered some challenges in reproducing the results from your paper.

	SA	RA	VA
(paper) baseline	84.90%	68.10%	1.3%
(paper) grafting (50%)	59.39%	45.77%	38.70%
(reproduced) baseline	84.97%	72.26%	1.5%
(reproduced) grafting (50%)	67.61%	54.61%	39.00%

I have a few questions that may help clarify the process:
Q1. Could you explain how the instability score is calculated? do you use the first 1000 samples as in your VA measurement? are incorrect or attacked samples included in this calculation?

Q2. At what stage are unstable neurons identified - before or after branch and bound?

Q3. If available, would you be able to share any source code, masking or weights related to these aspects?

Thank you for your time and assistance.

Warm regards,