On Explaining Neural Network Robustness with Activation Path

<p>ICLR submission. This work is supported in part by the Huawei Technologies Co., Ltd under Grant HIRP2019041002010, the Horizon<br> Europe: European Lighthouse on Secure and Safe AI – ELSA, the UK EPSRC under Grant EP/P009727/1, and the<br> Leverhulme Trust under Grant RF-2019-492 </p>