My conclusion now is GAN is useless, it does something probably mine does but by the wrong way. Reason is simple, while odd sentences may seem barely recognizable, the ex. odd word rearrangement is down-voted AND merged with other votes/weights, so it is useful accordingly, not overused. So I can't find any thing it does....GAN attacks weaker areas of the model more, there's nothing you can do here but update the ALL the model based on data samples, what can a 2nd model of the data do using the idea of "GAN"? Feed model a model? I give up for now here.
Ok let's try more, here's this, there's hope. https://developers.google.com/machine-learning/gan/generator some new information here full: https://developers.google.com/machine-learning/gan processing now, give me time I think this is going to do it, bets are on *So the generator is sending inputs to the discriminator binary CLASSIFIER along with real inputs too. During gen training, it uses the backprop loss that travels from disc output layer (that has 2 nodes binary) backwards to gen to update only the gen when failed to fool the disc. During disc training the disc loss updates only the disc when it is fooled by misclassifying a real instance as fake or a fake instance as real. Gen needs random input apparently. Training stops once each model is 50% accurate. Avoid Mode Collapse by averaging future discriminators. Wasserstein Loss is: Critic Loss = [average critic score on real images] – [average critic score on fake images. Generator Loss = -[average critic score on fake images].* Do I get it now? Will think about it in bed, not yet but I feel the big click coming. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T686c39065965ef71-M3685194ed7ee34615a099410 Delivery options: https://agi.topicbox.com/groups/agi/subscription
