Did you catch them? What if you invented them? Thanks to machine learning, a developer has created his own Pokemon and is encouraging computer enthusiasts to do the same.
Machine learning, or machine learning in French, is a branch of artificial intelligence that focuses on using data and algorithms to mimic the way humans learn, gradually improving its accuracy.
Concretely, this means that by giving access to data to certain algorithms, we can teach them to “learn”. We can thus train this type of algorithm to create images based on a set of data, such as the famous site This Person Does Not Exist which generates the faces of people who do not exist on each new visit to the site.
Transform them all!
It is towards this last category of machine learning that Max Woolf, a regular at computer content generation, decided to turn around. He thus set himself the challenge of making a computer learn what a Pokémon looks like, in order to generate it himself procedurally. A representation of all the Pokémon was thus provided to the computer which was able to learn to recognize them, define their characteristics and then recreate them. Surprisingly, the result is quite convincing.
While some designs are a bit too convoluted to be true Pokémon, the majority of Pokémon created are believable enough to fool an uninitiated. Faced with the success of his personalized pocket creatures, Max Woolf decided to publish more results.
For the more nostalgic, Max Woolf has also developed an algorithm based only on the 151 Pokémon of the first generation.
Available on Github
However, the good quality of the results obtained must be put into perspective. Indeed, this type of algorithm creates a very large number of results, most of which contain significant visual errors, making them unusable. Max Woolf therefore probably had to sort the results in order to keep only the best. Either way, the result is still a lot of fun.
Those who wish can also get their hands dirty and try the algorithm for themselves. Max Woolf has indeed made public his algorithm on Github. If you feel up to it, you can download it at this address.