formal verification of neural networks

Formal verification for neural networks is the problem of, given a [[ Neural Network ]] implementing a function \(\mathbf{\hat{x}_n} = f(\mathbf{x_0})\), a bounded input domain \(\mathcal{C}\), and a property \(P\), proving

\[\mathbf{x_0} \in \mathcal{C},\quad \mathbf{\hat{x}_n} = f(\mathbf{x_0}) \Rightarrow P(\mathbf{\hat{x}_n})\]

Properties can be represented as Boolean formulas over linear equalities that constrain the network’s behaviour.

Examples for properties include local robustness and [[ global robustness ]].


Sources

  • [[ Branch and Bound for Piecewise Linear Neural Network Verification ]]

Notes mentioning this note


Here are all the notes in this garden, along with their links, visualized as a graph.