COLLECTOR’S CHOICE - DeepDream by Alexander Mordvintsev

This month celebrates the 10th anniversary of DeepDream, an important development in the history of AI-generated art. Introduced in May 2015 by Alexander Mordvintsev, a researcher and artist based in Zurich, DeepDream was one of the first widely recognized applications of neural networks for image generation. It played a major role in popularizing AI art, inspiring a wave of experimentation that continues among many artists today.

Just before DeepDream: 1000 classes #3, 2015/01

Just before DeepDream: 1000 classes #2, 2015/01

Mordvintsev began exploring computers at an early age, which later led him to study computer science and computer vision. In 2014, he joined Google, where he started working with neural networks. As Google supported independent research, he used this opportunity to experiment with these networks and to reverse-engineer image-trained models in order to better understand how they process and interpret visual data.

Starting in January 2015, Mordvintsev began experimenting with AlexNet, a widely used image classification model developed in 2012. During this period, he generated a number of images that were later grouped under the series Just Before DeepDream, which demonstrated the early visual characteristics of the technique. Although these images lacked high resolution and fine detail, they already indicated the potential of deep neural networks for image generation.

On the night of May 18, 2015, he was inspired to run an experiment. Instead of allowing the network to process images layer by layer, he interrupted the process and manipulated the mid-layers, coaxing the network to enhance and generate features from partial data. The experiment produced strange, dream-like visuals that resembled hallucinations. He shared the results on Google’s internal network, where colleagues quickly recognized their significance.

Sunset, 2015/05

The program became known as DeepDream, and it is characterized by highly detailed textures and a surreal aesthetic. This visual style is similar to the psychedelic art movement, which emerged in the 1960s in connection with counterculture and often featured swirling patterns and distorted images. An interesting aspect of DeepDream’s early outputs was the frequent appearance of recurring visual motifs, especially dog and cat faces and eyes. This was due to biases in the ImageNet dataset used for training, which included a large number of dog categories.

DeepDream was released as open-source software in July 2015. Mordvintsev wrote a blog post about the method with Christopher Olah and Mike Tyka on Google’s research site. The publication and the images got wide attention online and were featured in many articles. A growing number of users started using DeepDream through websites and apps, which helped more people learn about machine learning and neural networks.

Father Cat, 2015/05

DeepDream is looking for patterns everywhere #2, 2015/05

DeepDream's public release also influenced several artists to explore neural networks. Mike Tyka, who had worked at Google, began creating art with DeepDream. Mario Klingemann also experimented with the medium, especially during his time at Google’s Machine Learning Residency. In 2016, the Gray Area Foundation for the Arts in San Francisco organized DeepDream: The Art of Neural Networks, one of the first exhibitions focused on artworks made using neural networks. The show included works by Alexander Mordvintsev, Mike Tyka, Memo Akten, James "Pouff" Roberts, and others.

DeepDream demonstrated how machine learning could be used to generate images and contributed to broader public interest in neural networks.

Many of Alexander Mordvintsev’s DeepDream works, created before the program was open-sourced, are now part of prominent collections such as Jediwolf, @BlueBeam9611, and more.

DeepDream: The Art of Neural Networks, Exhibition by Gray Area Foundation for the Arts, San Francisco, 2016, Source: grayarea.org