Multiscale Variational Autoencoder

Building a multiscale variational autoencoder (m-vae).

It is similar to a wavelet decomposition with a learnable encoding in the middle.

It creates different scale representation of an image and then encodes it into z-domain.

The downscaled versions are used to learn high level features and the higher level versions encode minute details.

The model allows to create as much as log_2(input_size) numbers of levels.

I intend to use this m-vae as a building block for Classifiers, Fuzzers, Anomaly Detection and more.

Tasks

• Build basic model
• Abstract Encoding and Decoding
• CIFAR10 notebook
• MNIST notebook
• Generator model
• Classifier model
• Anomaly detection

CIFAR10 Autoencoder

Trained on the CIFAR10 dataset to recreate images for 10 epochs.

Original / Recreation comparison

Cifar10 - Epoch 1

Cifar10 - Epoch 150