DeepDream | TensorFlow Core

Abstraction: Visualizing neural network patterns via gradient ascent on images

Key points:

• DeepDream amplifies patterns a neural network sees by maximizing layer activations via gradient ascent (not descent), coined "Inceptionism" by Alexander Mordvintsev at Google in 2015
• Uses InceptionV3 pre-trained on ImageNet; lower layers (mixed3) produce textures/edges, deeper layers (mixed5+) generate complex features like eyes or faces
• Loss is the mean activation of selected inception "mixed" layers, normalized so larger layers do not dominate
• Multi-octave processing (OCTAVE_SCALE=1.30, 5 octaves) reduces noise and builds progressively finer detail across image resolutions
• Tiled gradient computation splits large images into 512-pixel tiles with random shifts to prevent visible seam artifacts and reduce memory usage
• TensorFlow Lucid library extends DeepDream ideas for broader neural network interpretability research

Connections: Tensorflow · Google · Neural Network Visualization · Deep Learning · Convolutional Neural Networks · Gradient Ascent

Source: https://www.tensorflow.org/tutorials/generative/deepdream