Google's DeepDream was the 2015 computer-vision project that visualized patterns found by neural networks.
Google engineer Alexander Mordvintsev and colleagues developed the technique while studying the Inception image-classification network. Instead of asking the network to identify an existing image, DeepDream modified an image so that selected internal features became stronger. The resulting pictures often contained repeated eyes, animal shapes and surreal textures.
The method demonstrated that neural networks do not merely produce final labels. Their intermediate layers also encode visual patterns at different levels of abstraction. By amplifying those patterns, researchers could make some of the network's internal responses visible to people.
DeepDream is often confused with a conventional image generator or with Google's Inception network itself. Inception was the recognition model used in the original work; DeepDream was the visualization technique. Its public release helped popularize neural-network art and gave the wider public an unusual glimpse into how computer-vision systems interpret images.