Intriguing properties of neural networks – https://arxiv.org/pdf/1312.6199.pdf
- It suggests that it is the space, rather than the individual units, that contains the semantic information in the high layers of neural networks.
- It suggests that the learnt input-output mappings by deep neural networks are fairly discontinuous. That means it is easy to fool a well trained neural networks.
See also in https://arxiv.org/pdf/1412.6572.pdf where the authors show it is easy to generate adversarial examples which are close to the original ones but are misclassified by neural networks.