Paper Review: Intriguing Properties of Neural Networks

Alex Egg,

Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, Rob Fergus, ICLR, 2014


The paper introduces two key properties of deep neural networks:


The authors propose a way to make networks more robust to small perturbations by training them with adversarial examples in an adaptive manner, i.e. keep changing the pool of adversarial examples during training. In this regard, they draw a connection with hard-negative mining, and a network trained with adversarial examples performs better than others.

Formal description of how to generate adversarial examples and mathematical analysis of a network’s stability to perturbations are useful studies.

Two images that are visually indistinguishable to humans but classified differently by the network is indeed an intriguing observation.

This paper seems like the precurser to GANs, especially given its 5th author Goodfellow.

Permalink: paper-review-intriguing-properties-of-neural-networks


Last edited by Alex Egg, 2017-11-09 06:29:07
View Revision History