Neural synthesis studies

Summary Background Gallery How it works Tutorial + code

Neural synthesis of images and videos by stimulating channels in convolutional networks

References & background research

This technique grew out of deep learning research from 2012-2014 which aimed to improve our ability to interpret the activations and learned representations of convolutional networks, computational models which were rapidly gaining success in computer vision in that period.

Simonyan et al

Google cat thing ICML 2012 Quoc Le etc - https://static.googleusercontent.com/media/research.google.com/en//archive/unsupervised_icml2012.pdf

In 2013, Zeiler and Fergus

In 2014, Simonyan et al

In 2015, Mordvintsev, Olah, and Tyka

The works here are particularly indebted to Deepdream co-author Mike Tyka, who introduced … Gradient masks and iterative cropping methods