An outer-approximation guided optimization approach for constrained neural network inverse problems

This paper discusses an outer-approximation guided optimization method for constrained neural network inverse problems with rectified linear units. The refer to problem find the best set of input values a given trained in order produce predefined desired output presence constraints on values. analyzes characteristics optimal solutions activation units and proposes algorithm by exploiting their characteristics. proposed comprises primal dual phases. phase incorporates neighbor curvatures outer-approximations expedite process. identifies utilizes structure local convex regions improve convergence solution. At last, computation experiments demonstrate superiority compared projected gradient method.