In June 2015, Google researchers published “Inceptionism: Going Deeper into Neural Networks,” describing a technique soon nicknamed DeepDream. They took neural networks trained to classify images and ran them in reverse: instead of asking “what is in this picture,” they asked the network to amplify whatever it already thought it saw, producing surreal, hallucinatory images full of eyes, arches, and dog-like faces.
The method started from an existing image or random noise and iteratively adjusted the pixels to strengthen the features a chosen layer responded to. As the post explained, a network trained to recognize objects “contains quite a bit of the information needed to generate images too.” Different layers produced different effects - lower layers added strokes and ornamental patterns, higher layers conjured whole objects - which made DeepDream a window into what the network had actually learned.
The work straddled two purposes. As interpretability, it helped researchers see and correct what a network had latched onto. As art, it went viral: the psychedelic DeepDream aesthetic spread across the internet and was, for many people, the first time a neural network’s inner workings became something you could look at and share.
Why business readers should care: DeepDream marked the moment image-recognition systems were repurposed as image generators, an early hint of the generative-AI wave that would follow.