Digital Pareidolia

Medium: machine learning, video

About

This project uses machine learning to create an AI that identifies Christian imagery. The AI is fed thousands of conventional Christian images and then identifies what it deems as qualifying from video footage.

Description

The project will use a generative adversarial network (GAN) as its learning method to discover patterns in pre-existing images to identify them as outputs. The AI would be fed hand selected Christian images, not pareidolic images. A camera mounted on the head would capture video footage to be fed into the model. On regular intervals the model will inspect a given frame. The program would then identify and record what it identifies with timestamps and screenshots, tagging it with keywords. Technical requirements include: having knowledge of machine learning, choosing the best GAN for the size of the machine, acquiring compatible hardware for the camera input to computer, as well as creating a data set library large enough to feed the AI, and a name tagging convention.

Context

Pareidolia is the tendency to infer deliberate design or symbolism on a random or natural stimulus. Pareidolia often occurs with visual stimuli, when a specific arrangement of natural or accidental shapes can assume the form of a recognizable figure. Though a completely normal human function, even primal, we still do not know why we experience pareidolia. Studies have suggested that our brains have evolved to facilitate social interactions and helped our ancestors survive a hostile world; when looking for camouflaged predators amongst the brush, or foreign humans who might pose a threat, it is almost always better to be safe than sorry. This shapes the way we see and interact with the world around us. When we look up at a cloud, are we seeing the cloud, or our perception of what the cloud might be? We might see a dragon, a wagon, or an ancient cannon. Any number of visual things might manifest in our mind. Perhaps the most common form of pareidolia, is the perception of a divine or religious symbol in otherwise innocuous circumstances. There have been countless encounters with Jesus, burnt onto a piece of toast. Fishing a Cheeto out that looks like the Pope. Even in the vastness of space, a photograph captured by the Hubble Space Research Institute found many devout seeing Jesus kneeling with his hands in prayer, dubbing it the “Jesus Nebula.”

Technology is quickly becoming more and more “human.” As technology develops, and artificial intelligence gains more human characteristics, robots will perform many tasks more quickly, accurately, and efficiently than a human would ever be capable of. In which case, why not replicate a digital form of religious pareidolia. The purpose of this AI would be to identify and catalog instances of pareidolia that could be misconstrued as divine intervention. This project is also seeking to exemplify and highlight how human biases shape emerging technologies, influenced by different communities of people for numerous different purposes.

Data Sets

To the right are some examples of the labeled data set for training that would be fed into the program. This data would be the blueprint the model will use to find patterns.

The training data set would include selected pictures, specifically paintings and photographs of each category. The training set does not include pareidolic images.

Subcategories such as “crucifixion” and “cross” are meant to see how detail oriented the model can be.

Output

A camera mounted to a persons head records the footage as single image frames to scan. The wearer goes about their normal routine while the camera captures all of the memory.

Approved: Although just a tree, the model recognizes the pattern and with over 50% confidence identifies a cross.

Unidentifiable: This tree does not resemble any patterns the model has been trained to tag.

The level of confidence will be monitored to tweak thresholds for certain characteristics. This is important to ensure the model continues to “see” things that aren’t actually there.

The tagged images are then collected into their own folder at the end of the day. Users can look back at the documentation and view all of the tagged images.

Speculative Renders

Speculative renders of how the program identifies and labels pareidolia at the The grid overlay implies the pattern-seeking behavior of the model to make the most appropriate decision.