Generative Deep Learning with Variational Autoencoders

SPEAKER:
Alex Kalinin currently leads AI/Machine Learning at Meta. Previously, he developed smart home software at Home.ai using computer vision and deep learning. At Yahoo he led development of Big Data user acquisitions systems for Yahoo Games business. Alex holds MS in Physics, and published several papers on Image Recognition and Pattern detection.

LinkedIn: https://www.linkedin.com/in/alexkalinin/

SYNOPSIS:
In this workshop we will learn about generative models, and how they differ from discriminative models, e.g. classifiers. We’ll review the concepts of encoders and decoders, and learn the design of the variational autoencoders (VAE) - a type of neural networks used to generate images.

In the hands-on part we’ll build autoencoder and variational autoencoder networks, and see how they can be used to generate new images.

PRE-REQUISITES:

1. Working knowledge of Python.
2. Familiarity with convolutional neural networks. For an intro to convnets you can watch this lecture by Andrej Karpathy: https://youtu.be/LxfUGhug-iQ.
3. We will use Kaggle for the workshop. Make sure you have or create an account at Kaggle. Make sure you have at least 4 hours of GPU time remaining. Typical quota is 40 hours / week.

AGENDA:
(target time: 2 hours)

1. Discriminative models vs generative Models. The design of the autoencoder. Encoder and Decoders.
2. Hands-on:
a) Design and train an autoencoder to generate images of cylinders.
3. Variational autoencoders. Kullback-Leibler divergence.
4. Hands-on:
a) Train a variational autoencoder network, use it to generate new images.
b) Use VAE trained on CelebA dataset to generate new faces.