Hmmm.. Note this paper on current deep neural net vision systems and their pathologies... related to Facebook's DeepFace system...
http://arxiv.org/pdf/1312.6199v4.pdf It's an interesting observation, though certainly not at all surprising given the complex/chaotic nature of NNs as dynamical systems... One thing that is unclear from the paper is whether the particular phenomenon under discussion is specific to NNs trained via gradient descent and its variations, as opposed to also applying to NNs trained by other methods like, say, evolutionary learning... I am vaguely reminded of the literature on "deceptive problems" for GAs, by the way.... Those are a different thing, but what's interesting is that the problems proved deceptive for GAs are not deceptive for Bayesian Optimization Algorithms. Similarly, problem instances that are adversarial for NNs trained via gradient descent, may not be adversarial for NNs or similar networks that are learned in a different way... An interesting question seems to be: does the 3% error rate that existing top-grade NNs have on face recognition, occur on **different cases** than the 3% error rate humans have on face recognition? If the NNs and humans tend to screw up in the same cases in real image corpora, this suggests that this phenomenon of adversarial examples is a mathematical quirk that may not matter very much in real life.... If the NNs tend to screw up on different cases than people, then one would want to investigate whether the cases where the NNs screw up are correlated w/ these adversarial examples somehow... DeSTIN, the deep learning vision system we're playing with for OpenCog, is sufficiently different from the networks studied in this paper, that I wouldn't be confident the phenomenon noted applies to DeSTIN.... -- Ben ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
