Adaptive Discriminator Augmentation

 This video from DTC this week caught my eye.  It's about a technique to train GAN systems with limited data. By utilizing data augmentation to artificially expand the range of inout data used by the model for training.  

The discussion on 'leaking' of augmentations into the Generator probability  model is interesting.  Is it relevant to other recent unsupervised learning techniques?


Here's some more specific info on what the video is talking about.

Here's a link to the paper on Adaptive Discriminator Augmentation for GANs, titled 'Training Generative Adversarial networks with Limited Data'.

Here's an analysis of the technique described in the paper.


Observations

The 'mean of the system output representation' image kind of makes you wonder what this system is really learning?  Note how the output images are always so closely spatially matched up.  Just saying?  

Comments

Popular posts from this blog

CycleGAN: a GAN architecture for learning unpaired image to image transformations

Pix2Pix: a GAN architecture for image to image transformation

Smart Fabrics