Seriously here, GANs sound hogwash to me, I'm feeling it is just doing something else. You can't do backprop on backprop like this.....
I now found a even deeper article, that finally explains something clearer!!, using overly loads of text and algebra and explanations not even helpful >:\ https://towardsdatascience.com/understanding-generative-adversarial-networks-gans-cd6e4651a29 *"The brilliant idea that rules GANs consists in replacing this direct comparison by an indirect one that takes the form of a downstream task over these two distributions. The training of the generative network is then done with respect to this task such that it forces the generated distribution to get closer and closer to the true distribution."* After this quote is more such interestingness. It seems he says that we get an accurate model by, um, looking at another model at the same level of accuracy, erm, and, erm, um, but one such that is trying to maximize the error. decent explanatory video: https://www.youtube.com/watch?v=6v7lJHFaZZ4 ok....so....none you apes are gonna explain it better than that i bet.....so hmm, let me see, CONCLUSION: Directly updating a model causes it to predict better, very very straightforward. Indirect updating updates your accuracy model by being told from a 2nd model trying to model all that is wrong but falls within the other's model so to fool it. And vice versa, until converge to Nash Equilibrium. Hmm. Makes much more sense. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T71d305816eb449ff-M6d7a1796695b08f6fdaea54a Delivery options: https://agi.topicbox.com/groups/agi/subscription
