ABSTRACT
In this section we survey the iterative inversion of neural networks and its applications, and we discuss its implementation using gradient descent optimization. Inversion is useful for analyzing already trained neural networks, for example, finding false positive and false negative cases and answering related ‘what-if’ questions. Another group of applications addresses the reformulation of knowledge stored in neural networks, for example, compiling transition knowledge into control knowledge (model-based predictive control). Among the applications that will be discussed are inverse kinematics, active learning and reinforcement learning. At the end of this section, the more general case of constrained solution spaces is discussed.
