Download any 1000 or more ‘appropriate’ and publicly available images from the web. This could be part of a data set or something specific that you picked up or are interested in.
Apply the tSNE library to these images and construct low-dimensional embeddings for the images. Use these embeddings to then:
a) Start at any random image in the data set
b) Sequentially chain the next image to the previous image using a scoring function/probability based on the tSNE embedding. So you want to chain the most similar image to the current one and so on. Choose a frame rate that is appropriate to convert this chain of images into a video. Your video shouldn’t be more than 3 minutes long.