Understanding and mitigating exploding inverses in invertible neural networks