Generative Replay as Pseudo Rehearsal

Hello all,
In Generative Replay we need a Generator (Generator in GAN or Decoder in VAE). But for both of them, we need to sample from a ‘Distribution’ to get a latent vector to generate new data. But this ‘Distribution’ of the past data can be approximated only if we save all the past data or at least the ‘exemplars’. This again falls in the category of storing exemplars for replay, with an additional Generator model to generate pseudo data. Kindly correct me if I have got it wrong!

1 Like

Hi @Mochitha! You don’t need exemplars, in some cases just a couple of parameters.
In this paper @ggraffieti summarizes it very well: https://arxiv.org/pdf/2204.05842.pdf