We present a Lewis signaling-based collaborative ''guessing game'' that is capable of learning the joint distribution of labeled images. We adopt a VAE-like formalism to design this Lewis Signaling Recurrent Variational Encoder-Decoder Network (LS-RVED). The LS-RVED is a recurrent probabilistic encoder-decoder network, where at each time step, the encoder observes a portion from the labeled image and encodes that onto a latent variable (a signal). Given this signal, the decoder network tries to generate/''guess'' the labeled image. This process over time is used to train the encoder and decoder to generate effective labeled images.
The present version is for MNIST dataset. Shortly, we will provide for other datasets like: PCAM, Chest-Xray-14, FIRE, HAM10000 from the medical domain, and CIFAR 10, LSUN, ImageNet
run the train.ipython notebook