Machine Learning has come of age. Modern frameworks allow you to experiment quickly and see if a certain ideas has value. Here are some of those ideas worked out to various degrees of completeness
Gramzoom uses the key element of the Style Tranfer Algorithm to create an infinitely zoomable movie from any image. It works best with images that are self similar, but anything will really do. Can cats look evil? It might just break the Internet.
Using Recurrent Neural Networks to generate Icons and Hieroglyphs. The icons are encoded in a machine learnable format.
An implementation of a Spotify-like song radio based on Word2Vec. Based on a large set of crawled playlists and using those playlists as sentence equivalents. The Word2Vec algorithm then produces a vector per song.
Use a pretrained image classifier to power a reverse image search engine like the Tineye or GoogleImage search. Returns only from a rather select set of images.
An implementation of deep dreaming where the network is restricted to just black and white and starts with a blob in the middle. It forces it to draw from the center and create ink like patterns.
World map of uses Word2Vec to color a world map based on the distance between words and the names of countries.Country names are an interesting way to geocode the semantic values of words though a bit noisy.
Create universal numbers by comparing the edit distance between all numbers from wikitravel's phrasebooks and for each picking the 'median' one. As a side effect, create a tree an evolutionary tree.
Using the Artistic Style algorithm to restyle pictures of museums in the style of their most famous painting. This way you get immediately an idea of what to expect at these museums. Uses data from wikipedia, wikidata andthe wikistats. Uses Anish Athalye Tensorflow implementation
Mostly for the 2014 World Cup, with some adjustments to make it work for the European Championship in 2016,this model takes previous matches and tries to predict the future outcomes. The model is super simple andyet especially in 2016 did rather well.
This project uses the twitter api to automatically generate tweets based on what is commonly tweeted. It listens to the general tweetstream and captures fragments. By randomly recombining those fragments something that reads almost like real tweets appears. It also gives an interesting insight into what the average user uses twitter for.