eth-sri / mn-bab

[ICLR 2022] Complete Verification via Multi-Neuron Relaxation Guided Branch-and-Bound
https://www.sri.inf.ethz.ch/publications/ferrari2022complete
7 stars 6 forks source link

Confusion about formulation of multi-neuron constraints #7

Closed JacksonZyy closed 8 months ago

JacksonZyy commented 1 year ago

Hi,

I read your MB-BAB paper with great interest. And I noticed that you formulate multi-neuron constraints as follows: image I wonder why the constant term is encoded as -p(i) instead of just p(i)? Does it mean that the elements in p(i) are always non-negative? I would really appreciate your clarification. Thank you!

JacksonZyy commented 8 months ago

Answered in ERAN github