Suppose we have a neural network (such as google uses for instance) which detects an object in images, which could be a cat or car. Suppose that it is instead an alien artifact that we dont have pictures of. What methods are there given this neural network to produce an image of the object it detects in images?

  • $\begingroup$ Your question is “given our network X : images -> classes, and some class c, can we construct an image I such that X(I)=c?” And the first answer is “yes just do gradient descent to optimize I so that X(I) gets as close as possible to c” but this gives an image that looks random. One can add as a goal that the image looks normal (ie that it follows certain statistical properties that photographs tend to have) and the result is something like the “deep dreams” one can find research about. $\endgroup$ – Dan Robertson Jun 6 '18 at 20:56
  • $\begingroup$ @DanRobertson, nice answer! Can I encourage you to write that in the 'answer' block instead of as a comment, so we can upvote it, and so the question will be marked as answered? $\endgroup$ – D.W. Jun 6 '18 at 23:50
  • $\begingroup$ Read about en.wikipedia.org/wiki/Autoencoder and en.wikipedia.org/wiki/Generative_adversarial_network. $\endgroup$ – quant_dev Jul 20 '18 at 20:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.