When Machine Learning Meets Pokémon

Fun experiments with Tensorflow

With all the talk about machine learning, it’s sad that most resources out there are quite daunting for someone who doesn’t have a computer science background. How can we make machine learning more fun?

Can a machine recognize a Pokémon?

Tensorflow (Google’s open source machine learning library) has an image recognition API that outputs its understanding of the image. The score indicates the level of confidence. The original results look like this:

I made it less intimidating in the following examples (inspired by Bastien Petit’s Trying TensorFlow’s image recognition script).

After some experimenting, I realized that Tensorflow doesn’t do much with 2D Pokémon pictures. It just thinks they’re about comic books, but doesn’t get into the details. So I grabbed some screenshots of these cool 3D Pokémon renderings from the official commercials . Now Tensorflow can see more clearly.

Let’s check out Charmander:

King penguin? I think their body shapes actually are similar. 😅

Here comes our beloved Pikachu:

I didn’t even know what an axolotl is. 😮

How about some badass ones? Gyarados：

Triceratops? I can see the facial bone structure. 😃

Dark Mewtwo：

Well, technically it is flying. 😄

Let’s end with Snorlax:

Piggy bank 🐷 + killer whale 🐋 ? I like that combo!

Why can’t we create more tutorials and resources like this to make machine learning more approachable and fun?