Can we somehow constrain BoundedModule.compute_bounds() lower bound to not be negative?

[Soothysay]

Hi, I have been playing around with the library with different datasets and found that BoundedModule.compute_bounds() with options to give both lb and ub always gives negative lb values. Is there any way we can constrain lb to be >0?

[shizhouxing]

Hi @Soothysay , if it gives a negative lower bound, it means that the verifier can't certify a better lower bound (or a postive lower bound) at this time. We can't add a hard constraint (lb > 0) if that's the property we want to certify.

To get better bounds, you may consider:

• Trying a stronger verifier. When you are calling compute_bounds, did you try method=CROWN-optimized? You may also try our complete verifier (https://github.com/Verified-Intelligence/alpha-beta-CROWN) which can further refine the bounds by branch-and-bound at a larger computational cost.
• Training a better model by certified robust training (e.g., https://github.com/shizhouxing/Fast-Certified-Robust-Training)

[Soothysay]

Thanks for the clarification. This helps!