aniketjivani / variational-autoencoders

Experiments with variational autoencoders in Julia

Variational Autoencoders

Reconstruction of MINST digits during VAE training using Convoluation/Deconvolution Neural Network, 50 latent dimensions, 2 epochs shown

About

A library using Julia's Flux Library to implement Variational Autoencoders

• main.jl - run model with MINST dataset, this will be dropped later
• Model.jl the basic Model, for now it's just a basic VAE
• Dataset.jl the interface

Open Questions

• Can KL-Divergence and reconstruction error be better balanced?
• Can VAE be used as a pure clustering method? How else would the latent space representation be useful?
• Is it possible (in Julia) to reconconstruct the reverse transformation (decoder), for a given encoder?
• Can VAE be used for columnar data with missing inputs?

References

Tutorial on VAE
Tensorflow VAE
Flux.jl
Flux VAE
Auto-Encoding Variational Bayes
Stochastic Backpropagation and Approximate Inference in Deep Generative Models

Figures

Conv/Deconv VAE during 4 epochs of MINST training. 10 latent dimensions. MINST digits choosen at random from test set.