Computation Graph Pruning to Keep the Optimal Path

Given a computation graph that “looks like a tree” (assume it is a DAG rooted at a root node), each node is a neural-network operator (e.g., conv, activation). Given an input (or input spec), you need...