YinYang-GAN

YinYang-GAN: Phase Lock + Constructionism + GAN + Cross-Modality + Iterative Inference

structure illustration:

\(x_i \in P_i, i=0,1,...,M;\) \(x_i\):sample, \(P\):pattern space, \(M\):number of spaces.

\(\hat{x}_i = G_i(x_i) = EC_i(DC_i(x_i));\) \(DC\): decoder, \(EC\): encoder, \(G\): generator.

\(D_i(x_i, \hat{x}_i) \in \{0,1\};\) \(\hat{x}_i\):generated sample, \(D_i\):discriminator.

\(z_i = DC_i(x_i);\) \(z_i\): latent code decoded from \(x_i\) with \(DC_i\).

\(z'_i = \sum_{j\neq{i}}{T_{ji}(T_{ij}(z_i))};\) \(T_{ji}\):Translator from \(j\) to \(i\).

\(\hat{x'}_i = EC_i(z'_i);\) \(z'_i\):combined latent code, \(\hat{x'}_i\):final output.

\(\frac{\partial{D_i(x_i, \hat{x'}_i)}}{\partial{z_i}}.\) Differential to optimize on \(z_i\).

Approach#1: training a AutoEncoder instead of training a unstable GAN.
Average Instance: for each given associated observation \(z_j(j\neq{i})\),　there is an dynamic average instance \(\bar{z}_i=T_{j\rightarrow{i}}(z_j)\).
For instance, let label of 'Lady' be an associated observation, the visual　compensation will be a slim body with long hair, it stands for the average　instance of 'Lady' in visual space.