SIFT is a visual experiment. How it works:

• A generator systematically creates every possible image transform in YCbCr frequency space
• Each transform is converted to an image
• A convolutional neural network evaluates each image to determine if it is interesting or not, i.e. whether it is more likely to come from a reference set of real images (the tiny image dataset) or the set of images generated by SIFT
• Candidate images are tweeted and stored to file

The question this project attempts to answer:

Using brute-force combinatorics, what is the proportion of meaningful images among possible images?

As of December 2017, SIFT has evaluated over 1.4e10 images; many have been interesting but none have passed the final test: appearing to be a photograph to the average human. The system currently evaluates approximately 300,000,000 images per day on average GPU hardware. At the present rate, it will take approximately 10^800 years to exhaustively search all possible images.

Built with Python, TensorFlow, and Kivy visualizations.