Nonintuitive characteristics and intrinsic blind spots
Szegedy et al., 21 December 2013 (arXiv / ICLR 2014). Christian Szegedy of Google with coauthors including Ilya Sutskever and Ian Goodfellow, in the December 2013 paper "Intriguing properties of neural networks" (ICLR 2014) that first reported adversarial examples in deep networks.
Intriguing properties of neural networks
More from Szegedy et al. · AI and science · AI and security · Before ChatGPT
Who else?
What it refers to: Deep neural networks learned by backpropagation, which can be made to misclassify an image through a hardly perceptible perturbation the authors term an adversarial example.
Other voices
- “The problem of accidents in machine learning systems” Amodei et al.,
- “These models somehow just generalize dramatically worse than people” Ilya Sutskever, Builder,
- “Pointillistic masses of blurry memory” Gary Marcus, Researcher,
- “Machine learning has become alchemy” Ali Rahimi,
- “Deep learning is going to be able to do everything” Geoffrey Hinton, Researcher,










