Google Asked a Neural Network to Make Its Patterns More Visible

1. Convolutional Network Origin

Google engineer Alexander Mordvintsev developed a computer vision program that utilized a deep convolutional neural network to find and enhance patterns in images. The underlying software originated in a network codenamed Inception, which was originally built for the ImageNet Large-Scale Visual Recognition Challenge. The system applied algorithmic pareidolia to deliberately overprocess photographs and create surreal, dream-like appearances. Google popularized the term dreaming to refer to the generation of images producing specific activations in trained networks.

The visualization tool helped researchers understand network classification tasks.

2. Reversal Procedure

While training classifiers normally involves adjusting network weights to match inputs, the dreaming process holds weights fixed while adjusting the input image. The procedure increases selected network activations by changing pixels in the input. This reversal procedure utilizes a one-to-many mapping process that is never entirely unambiguous. Reiterating the adjustment over existing imagery causes adjacent pixels to change until algorithmic pareidolia results.

The optimization resembles backpropagation applied directly to input pixels.

3. Psychedelic Aesthetics

The generated imagery frequently features recurring animal patterns, eyes, and architectural structures embedded within natural landscapes. Observers noted that the visual output strongly resembled LSD and psilocybin-induced hallucinations. Researchers suggested this similarity indicated a functional resemblance between artificial neural networks and particular layers of the biological visual cortex. Modifying internal noise levels in deep networks reflected how biological hallucinogens alter serotonin systems in the brain.

Internal model variations generated distinct changes in output images.

4. Open Source Release

Following internal experimentation, Google researchers published technical details and released the underlying code to the public on July 1, 2015. The code was provided as an IPython notebook based on the Caffe deep learning framework and required minimal external software dependencies. Users could supply custom images, choose specific network layers to enhance, adjust iteration counts, and control zoom parameters. The release sparked widespread internet interest among programmers and digital artists alike.

Developers were encouraged to tag shared creations with a specific hashtag.

5. Proliferation of Tools

After the open-source publication, a large number of web services, mobile applications, and desktop software tools appeared on the commercial market. These applications enabled general users to transform personal photographs into psychedelic compositions without configuring raw code. Artists and musicians adopted the visual style for creative projects, including music videos and digital exhibitions. Researchers also proposed that adding dreamed inputs to training sets could improve training times for abstractions.

The software became a popular medium for computer-generated art.

6. Virtual Reality Adaptations

Researchers subsequently expanded the application of the algorithm beyond static two-dimensional photographs into immersive virtual reality environments. In 2017, a Sussex University group created a Hallucination Machine by applying the algorithm to pre-recorded panoramic video. Participants explored these virtual environments to study subjective experiences resembling psychoactive states. Experimental trials demonstrated that the virtual reality counterparts produced phenomenological similarities to psychedelic states compared to control videos.

Subjects reported altered perceptual dynamics during the virtual exposures.

7. Neuroscientific Studies

Later scientific studies investigated neuroscientific evidence linking the software outputs to human brain activity during altered states. A 2021 study published in Entropy recorded electroencephalography signals of human participants viewing standard movie clips alongside algorithmically modified counterparts. Researchers discovered that the modified videos triggered higher signal entropy and increased functional connectivity between brain areas. These measured metrics matched established biomarkers of actual psychedelic experiences in human subjects.

Follow-up studies measured cognitive flexibility changes after exposure.

8. Documented Impact

Google released DeepDream code in 2015, allowing people to run the visualization procedure on their own images. The repeated enhancement of features could produce recognizable shapes from patterns the network emphasized. The published code and examples documented that transformation process. They did not record a computer literally sleeping or dreaming; the name belonged to the image-processing method the researchers had made available.

The code let other users reproduce the image transformations.

Sources

• Original narrative adaptation — CC BY-SA 4.0
• Wikipedia contributors — DeepDream (CC BY-SA 4.0)
• DeepDream - a code example for visualizing Neural Networks