Tuan Anh Le

Attend, Infer, Repeat

19 February 2018

Generative network

Here is a pseudocode for the generative network \(p_\theta(x \given z)\) where the observation \(x \in \mathbb R^{D \times D}\) is an image and \(z\) are all the latent variables in the execution trace. \(\theta\) contains parameters of various neural nets in the generative network.

where \(\mathrm{STN}^{-1}\) is an inverse Spatial Transformer Network, and the \(D_\theta\) a parametric function.

Inference Network

Here is a pseudocode for the inference network \(q_\phi(z \given x)\). \(\phi\) contains parameters of various neural nets in the inference network.

where \(\mathrm{STN}\) is a Spatial Transformer Network, and the LSTM cell \(R_\phi\) takes in an (input, hidden state) pair and outputs an (output, next hidden state) pair.

[back]