It's a neural network so chances are, we'll never know how it came up with it's decision. For CNNs we can look at the output of each layers to get a hint but the more diverse the architecture is (such as in this example) the harder that becomes.