This project is by David Bachmann (bacdavid). It is not published or affiliated.

Obtain the latent variables that contain the maximal information wrt. sample (mutual information). This work is inspired by the InfoGAN (Chen et al., https://arxiv.org/abs/1606.03657) where the mutual information between selected channels and the sample is maximized.

Fig. 1: Perturbation of the single InfomaxVAE-style latent variable. It can be observed that the lighting is addressed by this variable: In other words, the maximum information is contained in the knowledge about the lumination.

Fig. 2: All variables but the single InfomaxVAE-style latent variable are masked for the reconstruction. This is most likely close to what a regular autoencoder with a single latent variable would produce.

Fig. 3: The remining 99 VAE-style latent variables result in a much better reconstruction than just the one InfomaxVAE-style variable. However, the lumination is still better when including the single Infomax-style variable, which naturally is an important factor for the reconstruction.

Fig. 4: Reconstruction when including both types of latent variables.

Fig. 5: Original samples.

• Typical VAE network for the generator: Encoder - Sampler - Decoder
• Mainly convolutional layers for the encoder and de-convolutional layers for the decoder with kernel size 5x5 and strides of 2x2
• Batch Norm followed by ReLU after the (de-)convolution
• 64 - 128 - 256, 256 - 128 - 64 - 3 (RGB) feature maps for encoder and decoder, respectively

• VAE-style latent variables are denoted by z
• InfomaxVAE-style latent variables by c

The original VAE-loss is

mse(x, x_vae) + KL(p(z | x) || p(z)).

By adding the mutual information term, the following is obtained:
Please note that [...;...] denotes the concatenation operator

mse(x, x_vae) + KL(p([z;c] | x) || p([z;c])) - I(x; c) 
= mse(x, x_vae) + KL(p([z;c] | x) || p([z;c])) - KL(p(c | x) || p(c))
= mse(x, x_vae) + KL(p(z | x) || p(z)).

In other words, exclude the InfomaxVAE-style latent variables from the regularization term.

Simply open the file

train.py

and perform the required adjustments.