In particular, for traditional software testing, if we have an if statement(e.g., if( x < 0) ... else ...), it's easy to generate tests that cover both branches. However, it's hard to use neuron coverage to generate all possible inputs. In addition, DeepXplore requires users to provide input seeds and make minor changes to the input seeds to get the difference- inducing inputs, but how to pick these input seeds? For an image classification model, if the input seeds do not contain any cat images, how to find bugs that can only be caused by cat images?