Alexander Mordvintsev (Google Research) — DeepDream

2015

Artists: Alexander Mordvintsev (Google Research) · Year: 2015 · Medium: Convolutional neural network visualization technique — released publicly June 2015 · Substrate: GoogLeNet / Inception convolutional neural network; gradient-ascent algorithms for feature visualization · Place: Zurich, Switzerland · Institution: Google Research In June 2015, Google Research engineer Alexander Mordvintsev published a technique for visualizing what convolutional neural networks had learned by running gradient ascent on trained networks and amplifying features the network recognized. The resulting imagery — photographs warped into dog faces, eyes, and fractal forms that the network had learned to detect — was unlike anything the internet had seen. Within two weeks of the DeepDream release, the technique had been reimplemented by thousands of hobbyists, and Google's own images were reproduced across every major news site and social-media platform in the world. DeepDream is not an artwork in the strictest sense. It is a visualization technique that Google Research published partly for scientific communication and partly because the results were too striking not to share. But it is the moment that AI image-generation entered mass cultural awareness, six years before DALL-E 2 and Stable Diffusion. The aesthetic DeepDream established — the specific look of a neural network's learned-features hallucination — continues to influence AI-art aesthetics a decade later. The technique also demonstrated, with a clarity that no prior AI moment had achieved, that deep neural networks were not simply classifiers but had internal representations with their own specific visual vocabulary. The implications of this for both AI research and for art practice have been large. Mordvintsev continues to work at Google. The original DeepDream notebook is still available open-source. Lore Mordvintsev was working on neural-network interpretability when he developed the DeepDream visualization; the artistic applications were a secondary discovery. Google's June 17, 2015 blog post, Inceptionism: Going Deeper into Neural Networks, introduced the technique to a general audience. The psychedelic dog-face aesthetic became, briefly, a meme; it became, more lastingly, an influence on every subsequent AI artist's visual vocabulary. Mordvintsev has continued to publish on neural-network visualization through the subsequent decade. Sources • Google Research Blog — Inceptionism: Going Deeper into Neural Networks (June 17, 2015) • Mordvintsev, Alexander — subsequent Distill.pub essays on interpretability • GitHub — original DeepDream Jupyter notebook