Embeddings are one of the few elements of machine learning that I enjoy in the age of our AI overlords. So when I read Max Woolf’s latest blog post about Pokémon embeddings, I knew I was in for a treat. Using just JSON metadata and images, he built embeddings to see how well they calculated cosine similarities between Pokémon. The results were really good given the data at hand, as Max alludes to in his conclusion:
In all, this was a successful exploration of Pokémon data that even though it’s not perfect, the failures are also interesting. Embeddings encourage engineers to go full YOLO because it’s actually rewarding to do so! Yes, some of the specific Pokémon relationships were cherry-picked to highlight said successful exploration […]
There’s a lot to go on but it’s digestible at all levels. If you like the idea of this, you should also check out poke2vec, a model trained on 2.3GB of fan fiction.
Filed under: data machine learning Pokémon programming