Prior reading: What Are Formal Methods? | Linear Algebra Proofs of Optimality
The Core Question
Given a neural network $f$ and an input region $\mathcal{X}$, what is the set of all possible outputs $f(\mathcal{X})$?
If we can bound this set, we can prove properties like "no input in this region produces a harmful output."
Why This Is Hard
Neural networks are piecewise-linear (ReLU) or worse (sigmoid, softmax). The number of linear regions grows exponentially with depth. Exact reachability is NP-hard.
Linear Relaxations
The key insight: replace the hard non-linear problem with a tractable linear over-approximation.
- CROWN: Propagate linear bounds layer by layer
- Alpha-CROWN: Optimize the relaxation tightness
- Beta-CROWN: Add branch-and-bound for precision
- Alpha-beta-CROWN: Combine both — state of the art in VNN-COMP
What You Get
A sound over-approximation of the output set. If the bounded region is safe, the network is provably safe for those inputs. If it's not, you might have a real violation or a loose bound.
Practical Limitations
- Scales to networks with millions of parameters, but gets looser with depth
- Tightness depends heavily on network architecture
- Only handles properties expressible as input-output constraints
Connection to Safety
This is the workhorse of neural network verification. Every other formal method for NNs either uses reachability internally or competes with it.